Inside the world's most elite student hackathon – Full Documentary on Stanford Tree Hacks 2026

[00:00] Oh. [00:01] We did this so many of these trials. [00:04] Okay, so that's when he gives out and we [00:06] see that was maybe even 50 seconds [00:08] before he actually failed, our [00:09] prediction was able to stabilize uh and [00:11] estimate when he would drop it. Very [00:13] cool. [00:16] >> [music] [00:41] [music] [01:09] >> Welcome to TreeHacks, Stanford's annual [01:13] student hackathon. [01:15] This is the 12th year that they've done [01:18] it and it is 15,000 people who applied, [01:23] of whom only 1,000 [01:26] were able to get in. Stanford paid for [01:28] all of those students to fly out here to [01:31] Palo Alto, to their main campus, to [01:33] participate and compete for $500,000 [01:39] worth of prizes. [01:41] Lots of big AI companies are here. [01:44] Got Anthropic, got OpenAI, we've got a [01:48] whole lot of API vendor companies, a [01:51] whole lot of developer tools. This is [01:53] going to be an exciting hackathon. I'm [01:56] so thrilled you all are joining [01:58] Let's go check out what people are [01:59] building. [02:05] Tell me a little bit about yourself. [02:06] Yeah, I'm a junior here studying math [02:08] and CS. Awesome. Yeah. Good combo. Yeah. [02:11] Yeah, and [02:13] how did you get into the TreeHacks [02:16] event? [02:17] I applied. [02:18] >> both last 2 years. I think this is the [02:20] second year where Stanford students also [02:21] have to apply, so I applied like [02:23] everybody else. Yeah. Do you know what [02:25] the admission rate was for Stanford [02:26] students? I think they said it was [02:28] Stanford students I don't know, but I [02:30] know in total it was 15,000 applications [02:31] they took 1,000, so Did you keep looking [02:34] at your friend? Friend, come over here. [02:35] Any ideas of what you're going to build, [02:37] what kind of tools you're going to use, [02:39] anything like that? I presume you're on [02:40] a team together? We are, yeah. [02:42] >> We're still thinking. I mean, there's a [02:44] lot of cool hardware stuff we're still [02:45] looking at [02:47] or even other potential ideas we had, [02:48] but yeah, we're still in the like phase [02:50] of exploration. Sahaj. [02:53] Sahaj Patel, how's everything going with [02:54] you? [02:55] >> Sir, it's going good. I've got a team. [02:57] You got a team already? We just started [02:59] hacking. I got the team like around [03:01] three or four days before. I was just on [03:04] on Slack channel looking at these cool [03:05] guys, all from MIT, all from really [03:08] great universities. [03:10] Watching the repos, I mean, ooh, [03:12] those open source guys, they're really [03:14] strong. They're going to put a good [03:15] show. So, what brought you all here? Uh [03:18] well, this team, I'd say. Um I'm new [03:20] student from high school. Uh we did [03:22] robotics with each other and we decided [03:24] we want to apply to TreeHacks and have [03:25] some fun and actually build something [03:26] impactful. [03:27] You all three or all four of you are [03:30] among the elite few, the one in 15 [03:32] people who got into the hackathon. How [03:35] did you get in? Like, how did you [03:36] convince them to let you in? What What [03:38] do you think they saw in you? I think [03:39] having a lot of really strong past [03:41] experience and credentials and just [03:42] proving that you can really do a lot of [03:44] good stuff here. I think that's the main [03:45] thing that helped me and many other [03:47] people. I think even for me, like I [03:50] highlighted all my research and my past [03:51] experience and then I've won a few [03:53] hackathons before as well, like UC [03:55] Berkeley and all. So, I mentioned that, [03:57] so maybe some of that, I guess. Yeah, [03:59] I've always been interested in [04:00] hackathons. I'm part of like the [04:01] organizing team for LA Hacks at UCLA. [04:04] So, I just had really highlighted my [04:06] passion for hackathons, and I think that [04:08] helped me get in. [04:09] I think for me it was more about like [04:11] research experience, and then also like [04:13] doing the YC AI Startup School, [04:15] like showed my passion for going there. [04:17] What tracks and what prizes are you all [04:20] considering pursuing here? [04:21] >> Yeah, so for me like the way I like I [04:23] went to for hackathon prizes, and then [04:25] even won a few of those. I'd say the [04:26] main thing that I did is I looked at [04:28] prizes that were particularly [04:29] interesting. I feel like those that tend [04:31] to get a lot of value, like those ones [04:33] like they usually get a lot of people [04:34] from that, but I think those are also [04:35] ones where you have like the most room [04:36] to like really experiment and try a lot [04:38] of new things. So, usually I like to [04:39] find prizes that are really worth your [04:41] while. And often times these prizes can [04:42] sort of intersect. So, for example, you [04:43] could win the grand prize, but you could [04:44] also win the sponsor prize as well. So, [04:46] being able to target multiple tracks at [04:48] once, and not just like limiting [04:49] yourself like putting all your eggs in [04:50] one basket, I think that's also [04:51] something. Like being like pretty like [04:53] making sure your project is something [04:54] that can really span a bunch of [04:56] different prize tracks at once, so that [04:57] can really give you the best chance that [04:58] in my opinion. [05:01] So, we're building an a agentic [05:04] assisted benchmarking platform for [05:06] basically assisting researchers where [05:09] they spend a lot of time manually [05:11] benchmarking their algorithms. So, we [05:13] want to speed this up, which will [05:15] greatly help the scientific community. [05:17] Let's say you have a sorting algorithm, [05:18] right? Now, you devise a new sorting [05:21] algorithm that can sort any database in [05:24] like of one time. But like you want to [05:27] test its speed results like or let's say [05:30] it's you want to do a comparison on [05:33] equal hardware against the other [05:34] algorithms in the space, right? What I [05:36] thought was if you just gave the code [05:38] for your custom algorithm, a few [05:39] research papers, and a prompt describing [05:41] the problem you're solving [05:43] to an agent, [05:44] >> Yeah. what it could do is it could [05:46] implement those algorithms [05:48] compare the plots, run the analysis on [05:50] your own. Additionally, what we want to [05:52] give a choice is have a research agent [05:54] as well that could research more [05:56] algorithms in the space for you that you [05:57] might have not explored, implement them [05:59] as well, and run compare like run [06:01] comparison plots for you. Yeah, speaking [06:03] of quantum, how did you learn about [06:05] quantum computing in general? [06:06] >> Also, the first video I saw about [06:07] quantum computing was like a 1-hour [06:09] 30-minute video from free code camp [06:10] actually. That's that's what entered me [06:12] into quantum computing and actually uh [06:14] since then I've been diving deep. I'm [06:15] actually an incoming software [06:17] engineering intern at IBM Quantum this [06:18] summer. Oh, really? You're entering [06:20] And so, you're going to have access to [06:22] the coveted quantum computer that cost, [06:24] I don't know, hundreds of millions of [06:25] dollars to build and keep cold and [06:27] everything, right? Yeah. So, the first [06:29] first time I learned Qiskit and like any [06:31] quantum concept was through your video. [06:33] Awesome, man. Well, congratulations on [06:35] that internship. Uh-huh, thank you. And [06:37] you know, I'm just here for fun to hang [06:38] out with my friends and you know, build [06:39] something cool and hopefully win some [06:40] prizes. And I understand that you won a [06:42] prize at Cal Hacks [06:44] >> Yeah. which was uh a hackathon about 3 [06:46] months ago. [06:47] >> Yeah. Yeah, that was honestly actually [06:49] like my second hackathon. For Alex and [06:51] Vern they've been, you know, they've [06:52] been doing this for years, but [06:54] >> Really? but yeah. [06:55] That's great. So, you got a pretty [06:56] seasoned team. [06:57] >> Yeah, for sure. I mean, you know, we we [06:58] plan on winning something, you know, and [07:00] yeah, yeah, with the project that we're [07:01] building we're going to have like lots [07:02] of chances for multiple tracks, so. [07:05] Yeah, and I see there's a hardware [07:06] component. Can you talk a little bit [07:08] about what you're planning on building? [07:10] We are basically making a camera and [07:12] we're going to CAD up a chassis for it [07:15] that allows us to have a camera that [07:17] takes photos and we have like a little [07:18] printer as you can see over there. And [07:20] what we're doing is that we're basically [07:23] planning on taking images and then [07:25] running diffusion large language or just [07:27] diffusion models to do something with [07:28] the images. Like for instance, you can [07:30] take a photo, you can say, "Hey, make [07:32] this look like, you know, some certain [07:33] art style or whatnot." and then print it [07:35] out using the printer. And we basically [07:37] are going to be applying for multiple [07:38] tracks because since this idea is very [07:41] general with yeah, the camera that takes [07:42] a photo and it prints something, we can [07:44] do a lot of different things. Like for [07:45] instance, for the Visa track, which is [07:48] basically figuring out a way to do [07:49] smarter commerce, what we're doing is or [07:52] the idea that we have is that you can [07:53] take a photo of something and use like [07:55] the the Perplexity search thing to find [07:57] like the exact items and the prices of [07:59] it and printing out that. And for [08:01] another example for for the Perplexity [08:03] track, which basically wants you to use [08:05] their API for search, you know, you can [08:06] like take a photo of some, you know, [08:08] really anything and it'll just like find [08:10] all the relevant information. Kind of [08:11] like what Google already does, but you [08:13] know, you can actually print out the the [08:14] information. So, here I'm back again [08:17] with Mentor Mates at TreeHacks, one of [08:20] the biggest hackathons here in um [08:22] Silicon Valley. So, you know, being able [08:24] to come out here, seeing how people are [08:26] able to build really hopefully is able [08:28] to inspire folks in Hong Kong to also [08:31] learn how to build in hackathons as [08:32] well, too. Cuz ultimately, what I'm [08:34] trying to do is figure out how can we [08:35] democratize entrepreneurial education [08:37] and bring hackathons to more places that [08:40] people would otherwise not be able to [08:41] find them. You are Chinat, [08:43] uh and you [08:45] have been doing hackathons for a while. [08:47] You even run hackathons. [08:49] You run uh the Mentor Mates hackathon, [08:51] of course, in Hong Kong. [08:54] What is your stre- Are you participating [08:56] in this year? No, this year I'm actually [08:58] This is the third TreeHacks that I'm [09:00] back at. And this year, I am a mentor [09:03] and not a uh participant cuz I figured [09:05] that it might be helpful to also be on [09:07] the other side of the table to [09:08] understand what is it like for folks to [09:10] be able to actually build out their [09:11] projects. And I feel like in the past, [09:13] being a participant, you know, you [09:15] really are kind of siloed into thinking [09:17] about how do you build the best project [09:19] in the you know, in the scope of the 36 [09:21] hours. And then the question now is then [09:23] how do I actually figure out how other [09:25] people are building as well, too. And [09:27] being a mentor helps me to take a step [09:29] back and see how other people are [09:30] approaching hackathons, as well. What If [09:32] you had like just three pieces of advice [09:35] for people who are participating in a [09:37] hackathon like this one, [09:39] competing with a very kind of like elite [09:42] pool of like 15,000 applicants, 1,000 [09:46] participants that got in, [09:48] what would be the key tips you would [09:51] give them to succeed? [09:53] Don't overthink. Don't overthink. Don't [09:56] overthink. All right. I think the main [09:58] thing that I've seen a lot of folks, you [09:59] know, are unfortunately unable to do is [10:02] is get out of their heads and just [10:03] really figure out what they actually [10:04] want to build and get it to work. Having [10:06] that confidence of knowing that no [10:08] matter what happens, you're still a [10:09] winner and that at the end of the day, [10:12] you're competing not against anybody [10:14] else, but instead with yourself, right? [10:17] Uh in terms of practical advice, I think [10:18] that really think about what you're [10:20] coming in to this hackathon for. Are you [10:22] a missionary or a mercenary, right? If [10:24] you're a missionary, then just build [10:26] whatever you want to build that really [10:27] speaks to your heart. But if you're a [10:29] mercenary, then you want to really [10:30] figure out are you trying to get for [10:31] that internship? Are you trying to get [10:33] that specific prize, right? In that [10:35] case, who do you want to kind of [10:36] communicate with and have that [10:38] conversation with? Hi, Mom. If you're [10:40] watching free code camp, He he he and I [10:42] were fanboy. I was like, "Yeah, I was I [10:44] just want to say hi to Quincy. He came [10:45] to our hackathon. I wanted to say hi." [10:47] And I was just like, "Is that the guy [10:49] from the original videos?" And I'm like, [10:50] "Yeah, man." So, you tell me a little [10:52] bit about it. What are you doing here? [10:53] Mainly here just fool around with like [10:55] build something not like too technical, [10:57] but kind of fun. Like push push like [11:00] dystopian stuff, you know, like one [11:02] thing I'm thinking about is a paper [11:04] shredder that has a $100 bill above it [11:07] and if you like stop being productive, [11:09] it slowly feeds the $100 bill into the [11:11] paper shredder so you get back on task. [11:13] I think destruction of currency is like [11:15] a federal crime though. If it's our [11:17] project, [11:18] >> and art, it's not. [11:19] >> Yeah. Oh, okay. [11:20] I'm here with Josh on. [11:22] Uh they are building a local Yeah, yeah. [11:25] >> AI video editor. Yeah. So, basically we [11:28] are building FlowCut. It's uh [11:31] it's a platform which uses local uh [11:34] compute for performance and AI agent for [11:37] execution. So, cutting, clipping, video [11:40] generation, all of that will happen on [11:43] your local compute, but you can [11:45] configure it with any API of your choice [11:48] on or run your own local LLM as well. [11:52] It will work with those things and this [11:53] will save a lot of latency, cloud [11:56] compute, [11:58] and eventually be more scalable with [12:00] your own software. You want a drink? [12:03] Can I just get some milk? [12:07] I love milk. This is the Hewlett-Packard [12:09] garage, the garage where they started [12:12] the company. They preserved the company [12:14] just like if you see it on the show [12:15] Silicon Valley. [12:18] They They like preserved the garage. [12:20] It's the same kind of thing. [12:22] All right, we're here with Dylan, [12:24] who's got a team [12:26] of some fellow developers, and you all [12:29] are building What are you building? [12:30] Yeah, so we're we're building Iris, and [12:32] it's an AI assistant that works across [12:34] all of your Apple devices. So, currently [12:37] when you talk to Siri, it would only [12:39] like do something on your iPhone or it [12:40] would only do something on your Mac, but [12:41] we want to build an AI agent that's [12:43] unified across all of your Apple [12:44] devices. So, for example, you can draw a [12:46] diagram on your Mac, and that would do [12:48] something on your computer. So, like [12:49] computer use tasks or like invoke cloud [12:51] code. Or you can like have a math [12:54] problem on your computer screen and [12:55] write the equation on your MacBook, and [12:57] some hints would appear on your Mac [12:58] screen [12:59] on the side. And so, this is like [13:01] basically unifying all of our Apple [13:03] devices together for using AI agents. [13:06] What do the next 36 hours look like for [13:08] you in terms of the specific sequence of [13:11] things that you want to happen? [13:12] Yes, I think we want to start with the [13:14] demo of doing math on your computer. [13:18] You should be able to write on your [13:19] iPad, and then the cloud code or [13:22] something on your Mac should be able to [13:23] read that and then suggest you hints. [13:25] And then maybe you can write something [13:27] on your iPad next to you. Um so that's [13:29] like the first working demo that we [13:30] want. Um and then later on we'll do more [13:32] complicated stuff like working in large [13:33] code bases. Like maybe it'll draw a [13:34] diagram of your code base and you can [13:36] interact with that and say like oh no I [13:38] don't want this thing. Um I'm going to [13:40] cross that out and we want to try a new [13:41] API or some new other direction that we [13:43] want to do. Nicholas. Yeah. Uh this is [13:45] your team? Yeah. Tell me a little bit [13:47] about what you all are building. [13:49] All right. So the problem we're trying [13:50] to solve is drug costs. To create a new [13:52] drug costs $3 [13:55] and 20 years. But there's so many [13:56] shelves or there's so many drugs that [13:58] got shelved, right? They got dropped [14:00] because of business issues. Because like [14:01] the company went under or maybe they [14:03] couldn't find enough trial participants. [14:04] That's 70% of them. 70%? [14:06] >> 70% of dropped drugs or shelved drugs [14:08] are because of logistical concerns and [14:10] not due to the drug itself. [14:12] So what we're doing is we're automating [14:13] a method that can look at the drug or [14:15] look at any disease and be able to say [14:17] like is there a dropped drug that could [14:18] be useful in like preventing this? If [14:21] you look at Ozempic, if you look at [14:22] Viagra, they both started as something [14:24] completely different before they got [14:25] transformed to the current degree. And [14:27] we're making a method that automates all [14:28] of that. [14:29] So how do you automate it? A lot of [14:32] subagents. Um the way we're working [14:34] right now is when you run it you have an [14:36] agent that's going to do a literature [14:37] review. Right? You have different things [14:39] called knowledge graphs which are we [14:40] able to encode every disease and every [14:42] um drug as like a certain vector and we [14:44] can compare it to existing drugs in the [14:46] field and existing diseases and what we [14:47] know. So we can say how likely is this [14:50] drug going to affect this disease. Um [14:51] and we have all these agents [14:53] individually do these tasks. There's [14:54] seven of them. Then they're going to [14:55] come together and they're going to [14:56] debate. Say this agent believes it's [14:58] going to be these top three drugs. [14:59] Another agent thinks it's going to be [15:00] these ones. And we're creating a skeptic [15:02] agent who disagrees with everyone else. [15:04] So at the end they're going to talk and [15:06] they're going to come to a common [15:06] conclusion of what drug or what three [15:08] drugs are best and what are the [15:10] different like pathways to getting it [15:11] approved, right? Vibha, uh [15:13] congratulations on being one of the [15:15] elite few chosen to participate in [15:18] TreeHacks. Uh [15:20] how how did you get in? So I really [15:22] wanted to go this year and I was like [15:24] hoping to get in and I think like they [15:26] released the application and I was like [15:27] one of the first few like to apply. So, [15:30] I'm glad that I did and like I'm glad [15:32] that I'm here and just to see like so [15:34] many different things and so many like [15:36] different events. And I think the [15:39] highlight was like today I got to see [15:41] Sam from like Open AI and that was like [15:43] pretty cool. Mhm. So, yeah. [15:45] Very cool. Have you all decided on a [15:47] project idea yet? We're finally like [15:49] we're kind of deciding on it yet because [15:52] we have like a lot of different ideas, [15:54] but I think we're just trying to find [15:55] like a good niche that is like pretty [15:57] applicable in all of our lives. So, [16:00] yeah. I think that's what we're trying [16:01] to do. How are you planning to navigate [16:03] the next 36 hours? Like what if you had [16:05] to map it out by like 6-hour chunks or [16:07] something, what would the next few [16:09] stages look like for you? So, I think [16:11] first of all we want to really decide on [16:12] the idea within like hopefully less than [16:16] the next 6 hours. And I think the [16:18] beginning would be like deciding what [16:20] each person is going to do, like decide [16:22] what tasks and stuff. And hopefully like [16:25] within like next the next on Saturday [16:28] basically kind of finish out like the [16:30] basic outline of the app. And if there's [16:32] anything that we want to add on to it, I [16:34] like I think we can like decide on that. [16:36] And finally I think we have just like 9 [16:39] hours on Sunday to finalize our product. [16:42] So, yeah. Grace, how's everything going [16:44] with you so far? Yeah, it's going great. [16:47] So, tell me a little bit about um your [16:49] idea for what you're going to build. [16:51] >> Yeah. So, a lot of robotics companies [16:54] still probably manually train their [16:56] robots by collecting like like uh [16:59] thousands of footage. What we're [17:01] thinking of doing is kind of [17:02] crowdsourcing that information. So, [17:04] robotics companies can maybe like put [17:06] out a prompt and regular uh people can [17:10] upload like them doing the task and that [17:12] kind of just helps um like for a small [17:16] for a small fee. Yeah. Interesting. [17:19] What's your interest in robotics? Are [17:20] you studying it? Yeah. Um actually, my [17:23] friend here [17:25] Daryl, he interned at Tesla. So, that's [17:28] where he got the insight. [17:30] Yeah, how do you all know one another? [17:32] Uh I mean like we just met off of Slack [17:34] at TreeHacks last year. So, like we just [17:36] like Yeah, we messaged each other and [17:37] then we like made a team. And then like [17:39] yeah, we just came we came we became [17:40] friends afterwards. [17:42] So, I'm Joe. And then so, the plan for [17:44] this is to make a jazz chord progression [17:47] suggestion. So, you would have the user [17:49] play some chords like that. And then [17:50] ideally [17:52] the computer would listen to the chords [17:53] that you're playing [17:54] and then suggest what comes next. Okay. [17:56] Yes. And the way we're doing this is [17:58] just with some analysis with the physics [17:59] behind the overtone series and some of [18:02] like just a little signal processing [18:04] with the the sound that you're hearing. [18:05] Uh what are you planning to build over [18:07] the course of the weekend? Well, we hope [18:08] to have [18:09] a front end that has like a graph for [18:11] you to traverse. So, as you're playing, [18:13] you're you're going through from node to [18:14] node. And then it suggests for you [18:16] different options for what you have. And [18:18] then we hope to hook up the piano and [18:20] the guitar to to have the demo live. So, [18:22] when you're playing a chord, we can [18:24] traverse the graph with you. So, you [18:26] play a chord and then it's like [18:29] try this. Yeah. And then you just go to [18:31] that chord. Yeah. And so, it's like [18:32] you're kind of like just doing what it [18:34] tells you. Well, you have different [18:36] options to do it. And you can add some [18:37] embellishments while you're playing the [18:38] guitar. So, it's kind of like [18:41] just dramatically simplifying [18:43] which chords are good options from [18:45] there. So, it's kind of like a tree that [18:46] you're traversing. [18:47] >> Yeah, yeah, yeah, yeah. Okay, that makes [18:48] sense. [18:49] >> [music] [18:57] >> So, what we were trying to do is [18:58] integrate a virtual reality world using [19:03] Unity. And what we're going to try to do [19:05] is pair it with the Zoom track um by [19:08] using like a world foundation model to [19:10] create an environment where teachers can [19:11] interact with their students. Okay, what [19:13] is a world foundation model? So, it's [19:16] like a diffusion model, like that can [19:18] make images and videos, but it works in [19:20] like 3D space in real time. So, you can [19:22] prompt it. So, for example, if you're a [19:23] teacher and you say I want to be in a [19:24] classroom, you'll be in a classroom. [19:26] Awesome. And you're going to build this [19:27] in 36 hours? Uh hopefully using like [19:30] open-source models on the internet and [19:31] things like that. I'm Arya. I go to [19:33] Stanford. I'm a Stanford frosh. So, it [19:35] was really exciting that this is right [19:38] here at Stanford. Um I think what I [19:40] wrote about in my application was um [19:43] over the summer I did this project with [19:45] data mapping um where I mapped like [19:48] Chicago's digital divide. I'm super [19:50] interested in how like tech and um like [19:53] social impact how tech can help social [19:55] impact. And how how did you all get to [19:57] know one another? [19:59] Yeah, um I think that was really random. [20:01] I think we all just sort of like [20:03] messaged each other on Slack after [20:04] putting some introductions and then like [20:06] hey, you seem cool. I really liked how [20:09] you brought this sort of aspect to a [20:11] previous project and I think we just [20:13] kind of like went from there and had a [20:14] call and then vibed with each other. [20:16] Have you all been able to kind of decide [20:19] what type of project you want to build? [20:21] Yeah, I think we've got a general idea [20:24] of the space that we want to go into and [20:26] what technologies we want to use. I [20:28] think we're thinking about using Zoom a [20:30] lot as like a core component of our app. [20:33] And then definitely going more into that [20:36] idea of human flourishing, I guess. Like [20:38] so how can we support like first [20:39] responders or like operators, for [20:41] example, when they have like juggling a [20:43] lot of calls? [21:01] Hey Luna, got some good eats for you. [21:06] Hi, I'm Lori. Um I'm feeling pretty [21:08] good. I'm a little tired but I feel well [21:10] fed and I feel [21:13] I'm really happy because I feel closer [21:15] to my teammates now. Like we're just [21:17] hanging out normally and I didn't know [21:19] them before I came here. So I think [21:21] that's super cool. My name is Adam. I'm [21:23] feeling super hyped for TreeHacks. [21:26] We've had like a bunch of fun events and [21:27] then like bunch of stuff. So yeah, super [21:29] hyped. [21:30] >> Hi, I'm Yinsa. I'm super excited because [21:32] I didn't know any of them before coming [21:33] here and we've bonded a lot. We've had a [21:35] lot of good food together, seen the [21:36] llamas and super excited for how our [21:39] project will end up. I love Free Code [21:41] Camp. I learned HTML and CSS last summer [21:44] from Free Code Camp and I would not be [21:46] here if it weren't for Free Code Camp. [21:48] >> I followed the like how to code like a a [21:49] Discord bot like from scratch like on [21:51] Replit and in high school and I actually [21:54] used that project like on my common app. [21:56] Actually helped me get into college so. [21:58] >> How are you feeling about 15 hours in? [22:00] Very excited despite not having much [22:02] sleep over like the past 3 days. [22:04] But I don't know why I feel really [22:06] energized. Yeah, I feel like our team is [22:08] making good progress and we're really [22:10] enjoying working together. We met for [22:11] the Some of us met for the first time so [22:13] it's been fun working with everyone and [22:15] generating new ideas together. So hey, [22:17] hi. My name's Priya and I'm here [22:19] actually from Seattle area. I'm one of [22:21] the mentors. [22:23] So and this is Malika. We're both here [22:25] as mentors so. [22:29] We are absolutely impressed with the [22:32] number of the brilliance of students and [22:34] the ideas that they're coming up with [22:36] and the kind of questions they're asking [22:37] like I have to go back and figure out [22:39] okay, what is this? Right? So it's just [22:42] us AI and students together trying to [22:44] hack something out here and the energy [22:47] is amazing. The vibe is super cool and [22:51] the innovation is in the air pretty [22:53] much. [22:54] The the kind of ideas that they're [22:55] thinking about I love the social impact [22:58] angle that a whole lot of ideas have so [23:00] they're really thinking about solving [23:02] some of the big challenges in the world. [23:04] So, it's not just about, "Hey, I just [23:05] want to slap technical complexity into [23:07] it." But, they're really starting to [23:09] think about the angle is, "How do I [23:11] solve this problem in the world?" Right? [23:12] And that's what gets me really excited [23:14] about helping these kids out and just [23:16] trying to see if we can build something [23:18] cool together in this 2 and 1/2 3 days [23:20] we have here. So, yeah. [23:22] That's what I would say. [23:23] >> Right on. Would you like to add [23:25] something to that? Uh yeah, for sure. [23:26] Hi, I'm Malika. Uh I mean, it is not 18 [23:29] hours in for us because I woke up I and [23:33] I came here at 9:00 in the morning. So, [23:35] I'm very impressed by the tenacity of [23:36] these students and the how they're [23:39] continuing to go with high energy. Uh I [23:42] am [23:43] blown away by the innovation that the [23:45] students are thinking about and how [23:47] these projects are uh I wish I could [23:48] stay and watch all of them pitch their [23:50] projects. Is there any particular [23:52] project that you've met for where you've [23:53] been like, "Wow, this might win."? Like, [23:55] that's a hard question, right? Because [23:56] every project is in a different area. [24:00] Like, I was talking to a group of [24:01] students who are building a health care [24:04] agent that's going to follow up with [24:06] patients on recovery recovering from [24:07] surgical ops and uh just checking up on [24:10] them is like, "How's the recovery [24:11] going?" Because there's a lot of [24:13] relapses that happen. And having worked [24:16] on something like that within the [24:17] industry, I know that that's a big [24:19] problem. So, I thought like they had a [24:21] really cool problem statement to solve. [24:23] So, that was really cool one in terms of [24:25] like impact. There's this another one [24:27] that I absolutely loved from the point [24:28] of view of just technical complexity. [24:30] So, these people uh there's a group of [24:32] students they're using a camera uh [24:35] installed on like a flying robot, right? [24:37] And they're going to see if there are [24:39] any fires or any natural calamity long [24:41] horizon. And they are using computer [24:43] vision models and AI to figure out uh [24:46] what is the propensity of like how fast [24:48] it would spread. They're also figuring [24:50] out what might be the way to remediate [24:53] that. So, I was like, "Guys, you have [24:56] picked up a big one. But, if you can [24:58] solve this, I can see how it could be [25:00] used, you know, in disaster relief or [25:03] really solving some of the big problems [25:04] in the world. [25:05] >> There's There's two more that caught my [25:06] attention, and I will admit that I'm a [25:08] little biased because these were [25:09] students that we sort of helped with. [25:11] One was [25:13] powered boat that can collect trash in [25:16] the water as it goes and also try to [25:19] identify what pieces of trash can be [25:21] upcycled. [25:23] Those those kids are have been working [25:25] hard at it. The other one that I really [25:28] really was impressed by was this [25:31] system around [25:34] managing food banks. So, they have this [25:37] AI agent that connects with food banks [25:39] in the area, tries to take stock of [25:41] their inventory, and try to figure out [25:42] which [25:43] which types of food items they actually [25:45] want. And then they have a robotic [25:47] system, an intake system that accepts [25:50] donations either in kind or cash, and [25:52] then sorts them out and ships them to [25:54] the right food banks. And then they have [25:56] cameras inside the food bank where for [25:59] people who have donated stuff, they can [26:00] see a real-time feed of oh, that that [26:03] can of beans that I donated, this is [26:05] where it's going, and it's actually [26:06] going and helping someone. To be able to [26:08] give people that food feedback loop. [26:10] Yes, and then the reinforcement of what [26:12] value they have provided. [26:13] >> Our project for TreeHacks, [26:16] it's to solve like this medical [26:18] condition called aphasia. Sometimes when [26:21] people get like a stroke, it'll like [26:23] affect their ability to produce like [26:27] you know, speech. What we're basically [26:29] trying to do is create a device where [26:31] they can talk to you know, a microphone, [26:34] it's processed, and then we have like [26:37] AI try to predict what they're trying to [26:39] say. It'll give them some options on a [26:42] screen of what they it thinks they were [26:43] trying to say. [26:45] And they can just select that option and [26:47] it'll like talk for them basically. [26:49] All right, yeah. So, this This just like [26:50] a simple kind of web app demo just to [26:53] demonstrate that the code works. [26:55] Um [26:56] this like yellow button will correspond [26:58] to like the actual physical yellow [27:00] button here. Obviously, the design will [27:02] have these like really big tactile [27:04] buttons and then a mini [27:07] um screen that will like, you know, [27:10] put what's on here on here. The user [27:13] will press the yellow button, but it's [27:14] on the web app cells. And I'll say like [27:16] some jumbled phrase like, "Hi, my name [27:20] is Pierre hacking tomorrow when event" [27:23] something. And then it'll process [27:26] um what I said this using the Open AI [27:28] like Whisper [27:29] API. So, the input is [27:31] recorded was, "Hi, my name is Pierre [27:33] hacking tomorrow when event" something. [27:36] And it generated um three responses. So, [27:38] I'll click that. [27:40] It'll play it. [27:41] >> is Pierre. I'm hacking tomorrow. When is [27:43] the event or something? And it'll record [27:45] that response and then store it in the [27:47] database. And then I could do it again. [27:49] And we want to make sure that we could [27:51] use large language models not to just [27:53] like speak for you, but actually be [27:55] assistive. So, for those milder cases, [27:57] proposing words, for example, that allow [27:59] you the person to kind of find their way [28:02] to what they're wanting to say. Um of [28:04] course, but like for the more severe [28:05] versions, that's when like AI can help [28:07] up with that. [28:09] Yeah, so I'm Sheamus. I'm a master's [28:11] student actually in public policy here. [28:13] Uh so, I'm not really a computer [28:14] scientist, but what I do as kind of my [28:17] other passion besides public service is [28:19] game design. So, I'm really interested [28:22] in the ways that games can tell stories [28:25] that you can participate in, the ways [28:27] that they can kind of facilitate [28:28] connection between people, which as [28:30] we'll get into is a lot of kind of what [28:32] we're doing here with this project. And [28:33] so, you know, I put on my TreeHacks [28:35] application. I was like, "I haven't done [28:36] a hackathon before. I'm not a CS [28:38] student, but I'm a game designer. I've [28:40] made games before and, you know, I'd [28:41] like to come here and make a game." And [28:43] I guess the organizers were like, "Sure, [28:44] that sounds great." And so, [28:45] >> Awesome. that's what we're doing. I kind [28:47] of started developing games uh before [28:50] all of these AI tools were as powerful [28:52] as they are now. And I do wonder about [28:55] the effect that it's going to have on [28:57] people coming from kind of that liberal [28:59] arts background with the tools already [29:01] being so powerful that I feel like some [29:04] people might miss that important kind of [29:06] learning experience of figuring out how [29:08] to do the stuff on your own so that, you [29:10] know, when the AI agent messes up or [29:12] gets confused or stuff, you can actually [29:14] dive into the code yourself and figure [29:15] out what's going on. So, you know, I'm [29:17] I'm hopeful that the the great educators [29:19] that we have, you know, all across the [29:20] country will continue to to find ways to [29:23] make sure that people are [29:25] able to like collaborate with agents in [29:28] good ways and use the agents in ways [29:29] that actually help them understand what [29:31] they're doing cuz I feel like I would [29:33] get pretty overwhelmed if I hadn't, you [29:35] know, had that background in game [29:36] development and front end before I [29:38] started using all of the the AI agents [29:40] and tools. What are you doing exactly? [29:42] And what are you building and how are [29:44] you building it? We are building kind of [29:45] a daily [29:47] game kind of in the style of Wordle, but [29:49] the point is hopefully to connect people [29:52] with their friends. So, the way it works [29:55] is we have two kind of axes on like a [29:57] big graph or chart that are labeled [30:00] silly things. So, one might be like [30:02] muffin to pancake, another one might be [30:04] like Gimli to Legolas. And the idea is [30:07] that you place yourself and your friends [30:09] on this graph based on kind of like the [30:11] vibes of it. So, like where you whether [30:13] you think you're more of a muffin or [30:15] more of a pancake, whatever that means [30:17] to you. And you're trying to guess where [30:19] your friends will place themselves. So, [30:22] basically you get more points if you're [30:23] able to accurately figure out how your [30:26] friend placed themselves and place them [30:28] where where they placed themselves. And [30:29] similarly, you get points if your [30:31] friends kind of place you accurately. If [30:33] you're known well. [30:36] We're about maybe like 16, 17 hours into [30:39] the hackathon. [30:40] >> Right. [30:40] >> How are you feeling? Oh, I'm feeling [30:42] great. I mean, the food is wonderful. I [30:44] love the acai bowls. Acai bowls were [30:45] really great. And I have a really great [30:48] amazing team. We have made like a lot of [30:50] progress like All right. Tell me a [30:52] little bit about your project. Yeah, so [30:54] we [30:55] we recently um decided on a name. Our um [30:59] name of our project is called [31:00] Pigeonhole. The whole idea is that um [31:03] you know, AI usage is sometimes seen as [31:05] a bit of a problem for teachers because [31:07] it's difficult for them to kind of [31:09] manage um you know, what students are [31:12] kind of getting out of that. Um and so I [31:15] feel like it's really we have to kind of [31:17] have come to accept that like AI usage [31:18] is just going to be kind of part of our [31:20] daily lives. So we are creating this [31:22] tool that allows teachers to be able to [31:25] kind of dictate a little bit like what [31:27] um students are able to kind of get [31:29] access to through like a chatbot. And in [31:32] that way students are still learning [31:34] through like their usage of like AI, but [31:36] teachers will have a little bit more [31:37] control over that. [31:41] >> I'm Magnus. I'm a first-year PhD student [31:44] at Oxford. [31:45] Um [31:46] you asked what the vibes were. Uh [31:48] the vibes sort of start high and then [31:50] sort of a slow decline as you actually [31:51] have to do the work. [31:53] Um but then we've started getting some [31:54] results and some traction, so [31:56] it's on the high even though sleep is on [31:59] on on the low. So my name is Ray. I'm a [32:01] senior at NYU studying math and computer [32:04] science. [32:05] I really like this hackathon. I'm very [32:07] excited about the speakers. And yeah, a [32:10] lot of very like interesting people and [32:13] they are building very like uh [32:15] surprisingly good projects. And yeah, I [32:18] like meeting all those ambitious like uh [32:20] students. And yeah. Tell me a little bit [32:23] about what you all are building. [32:24] >> We find the like companies, they always [32:27] have like a they always spend extra time [32:29] and extra money on like off-boarding and [32:30] onboarding process. So we're trying to [32:33] like uh using the energetic AI stuff to [32:36] like accelerate the process. Yeah. The [32:38] key issue is that when someone leaves [32:40] the company often in a hurry a lot of [32:43] information or knowledge they have [32:44] leaves with them. It It goes and it's [32:46] It's hard to recover and people aren't [32:48] very motivated to [32:50] write nice handover docs and [32:52] often when they're leaving on [32:53] unfavorable terms. [32:55] So, we thought with AI being so [32:58] analytical and insightful these days, we [33:00] can make this process a whole bunch [33:01] easier. Usually it takes like 6 months [33:04] for a team to replace someone who's [33:06] leaving with a new hire. [33:08] And during that process like the work is [33:10] being stalled. It's like [33:12] productivity decrease. But essentially [33:14] like if we can keep like a digital [33:17] replicate of that person even after they [33:19] leave and then use that as like the best [33:21] onboarding tool or like just like a Q&A [33:24] for the rest of the team, then they can [33:25] keep work going. [33:33] There is a [33:34] kind of like bicyclist not following the [33:36] rules epidemic here on the Stanford [33:38] campus? [33:39] So, my roommate and I came up with a [33:42] kind of humorous solution which was to [33:44] build a annoying robot that follows [33:46] students who are biking without bike [33:48] lights at night [33:50] to tax them using computer vision and [33:52] follows them [33:53] and yells at them. Yeah, so my partner [33:55] Jay and I, we're building a what we call [33:58] magic table. It essentially uses like a [34:01] 2D gantry system and magnets to bring [34:04] objects to you. So, it's to help blind [34:07] people and disabled people who don't [34:09] necessarily know what's on their table. [34:11] They're able to talk in natural language [34:13] and have it just bring it straight to [34:15] you. How far are you along with the [34:16] prototype? Well, we we've 3D printed [34:18] most of the parts. I've been I'm [34:20] designing a scissor lift right now and [34:22] then we're going to reprint some other [34:23] parts as well. But so far we've got all [34:26] the hardware ready and we're just [34:27] assembling together and the code has [34:29] been He's He's working on of the code. [34:31] >> Maxwell Yes, Maxwell. You're an [34:33] organizer. [34:33] >> I am an organizer. [34:34] >> So, you thank you for helping organize [34:36] this hackathon. Yeah, well, tell me a [34:38] little bit about yourself. Yeah, I'm a [34:39] freshman here at Stanford like a lot of [34:41] people on the team. Um, I'm on the [34:43] logistics sub uh organizing sub team. [34:45] So, I mean, I guess kind of [34:48] all the like prep to get the event [34:50] going. I'm personally I worked on the [34:51] food. So, getting the caterers, [34:53] organizing the meals, and like snacks, [34:55] drinks for um like everybody. So, yeah, [34:58] it's lots of fun. You know, and a big [35:00] part is the builds. I don't know if [35:02] you've heard about them. [35:03] Um [35:04] that was a lot of work. We got like 3 [35:06] hours of sleep the night before, two [35:07] nights ago. So, [35:09] we're all a little tired, but it's very [35:10] fun. [35:12] Great time. Yeah. Yeah. [35:13] >> Yeah, well, what are you doing here in [35:14] the hardware lab? I was just walking by [35:17] um with one of my with my girlfriend [35:18] actually to see what she was working on. [35:19] >> That's cool. Is she participating in the [35:21] event as well? [35:21] >> Yes, she is. She's hacking. [35:22] >> Very cool. Um yeah. Should we interview [35:24] her about what she's building? [35:25] >> we should. Oh, we're like building an [35:28] organ transport kind of monitoring [35:30] system so that when you're on ground [35:32] level and you're like transporting [35:34] organs for donations, you can make sure [35:36] that they're not going to get damaged. [35:38] Okay, and what does the uh system look [35:39] like? Right now, it is a bunch of [35:42] sensors um hooked up to a front end that [35:44] moni- like connects several different [35:46] hospitals, and the goal is to collect a [35:49] lot of data uh correlated with like [35:52] where the nearest donation like centers [35:54] are, and if things go wrong, like how [35:56] you can adapt to them uh using different [35:58] routes. My understanding is you just got [36:00] here. You arrived like 18 hours late, [36:03] and now you've got this much time [36:06] to come up with an idea and build it. [36:08] Yeah, so basically my group, this is [36:10] Panos and I, um we came from a [36:12] sustainable global sustainability [36:14] engineering challenge yesterday. Um so, [36:16] we had the showcase for that, and we got [36:18] some sleep, and got here yeah, about an [36:20] hour ago, had some lunch, and now we're [36:21] starting to ideate on to as to like what [36:23] we can accomplish in 18 hours. What [36:26] broad strokes like ideas have you been [36:28] batting around? I'm in mechanical [36:30] engineering. [36:31] I'm in management science and [36:32] engineering. We've been trying to do [36:33] something hardware or something robotics [36:35] related. [36:36] So, we're just talking to the hardware [36:39] people over there seeing what kind of [36:40] components we can use. [36:42] Unfortunately, this year it seems like [36:44] there's a bit less [36:45] than in previous years. They have a DJI [36:48] Spark drone and some compute and the [36:52] Meta Quest goggles. So, we're thinking [36:54] like maybe maybe use the drone, build [36:56] something based on that with some like [36:59] computer vision system maybe. So, So, a [37:02] funny story. Like on the way from [37:03] Atlanta to San Francisco, I had no Wi-Fi [37:06] on the Frontier plane and I was bored, [37:08] right? So, I went on my downloads to see [37:10] if there's anything fun to do and I [37:11] think 2 years ago I downloaded a [37:13] tutorial from you guys on YouTube [37:15] on how to learn C. So, I was just [37:17] watching that C tutorial Watched it C on [37:19] the plane for 3 hours? [37:20] >> Yes, for 3 hours. Very cool. So, it got [37:22] me through the plane. Right on. What are [37:24] you all building so far? We're building [37:26] like an [37:28] EMG [37:29] signal to text device that can [37:32] essentially go through and get [37:34] like actual signal data, basically get [37:37] the signal data and then convert that [37:39] into like text so that we can pass that [37:40] into like a language model that can then [37:42] like autonomously do actions on your [37:44] browser. So, we're currently working on [37:47] getting our like browser automation to [37:49] work. We have like our hardware where [37:51] we've been able to collect a good amount [37:53] of data. So, we're going to see where [37:54] where that takes us. Basically, what [37:56] he's saying is we're using our mind to [37:58] control a computer without talking, [38:00] without typing. How does it work? Like [38:02] how does it capture the signals to begin [38:04] with? We're using like a muscle sensor. [38:06] We put put it like underneath your jaw [38:08] and then whenever you think like your [38:11] muscle moves somewhat and like that is [38:13] like signals that you can like record, [38:15] right? And then we take those signals [38:17] and then put it through like a model [38:18] that we can like analyze and [38:21] like determine like text from that what [38:23] you were thinking. So. [38:25] We're here at Robot Fight Club where [38:28] developers are testing their might with [38:31] these robots programming them to see who [38:34] can win. It's like BattleBots, the old [38:36] TV show. [38:40] >> [music] [38:47] [music] [38:55] [music] [39:01] >> So, we're working on a [39:04] voice AI centric platform that allows [39:08] nurses to get more specific reports [39:12] within the context of [39:14] uh nurse nurse [39:17] nurse patients kind of communication. [39:19] You're getting too much coarse or vague [39:22] data from the nurse. Yeah. The stuff she [39:25] inputs into the patient chart or [39:26] whatever data collection system. Right. [39:28] >> So, you are basically trying to like [39:31] infer what she actually means [39:33] he or she means by their notes. Yeah, [39:36] essentially. [39:37] >> AI to jump to make that inductive leap. [39:40] Yes, we're making that inductive leap to [39:44] not make the decision for the nurse [39:46] themselves, but rather guide them in the [39:48] direction of of actually getting that [39:51] more specific data. So, as an example, [39:54] um I have a patient that's within this [39:58] uh [39:58] that came to me and this patient just [40:01] underwent surgery. [40:03] And [40:04] what we could describe it as was the [40:05] patient was kind of drowsy. So, now the [40:08] AI leads me into asking what what is the [40:12] current level of consciousness? Like [40:14] what what So, now as a [40:16] nurse I'm thinking, "Okay, I need to [40:19] full talk about follow-up questions [40:22] relating to why was the the patient [40:25] drowsy? Drowsy is a very general [40:27] description. That would be very like [40:30] uninformative in the sense of a [40:32] generated report [40:34] within the for a doctor or anything like [40:36] that." So yeah, the AI tools are really [40:38] really good nowadays. And what we [40:40] thought would take us 36 hours got [40:42] completed in 6 hours. Wow. And after [40:46] that, we thought how do we make it [40:47] unique? So [40:49] we have thought of a universe where you [40:51] can train your own directors by giving [40:53] them your own style, your videos. They [40:55] learn and there's a marketplace of all [40:57] of these directors. [40:59] And these directors essentially critique [41:01] your video and then you can choose [41:04] multiple of these directors. [41:05] >> So a lot of people have been hyping up [41:07] what you've been building. A lot of [41:08] people have been telling me, "We got to [41:10] go check this out." So tell me a little [41:12] bit about what you've built so far. [41:13] >> My name is Ryan Nee. [41:14] Um I go to hackathons every single [41:16] weekend. My most recent hackathon [41:19] I actually went to Open AI's hackathon [41:20] and we won over there. Yeah, so what we [41:21] built Have you played Minecraft before? [41:23] >> Yeah. So do you do you know about 2b2t [41:25] by by chance? Not really. [41:26] >> No worries. So it's the first anarchy [41:28] server for humans to play games to play [41:30] Minecraft together. It's the It's the [41:31] first server. [41:32] And so what we are building we're trying [41:34] to make history at TreeHacks and so [41:35] we're building the very first anarchy [41:37] server for AI AI agents to play [41:38] together. And how we what we did is we [41:40] built like a harness for AI agents for [41:43] on Openclaw. We just built we just [41:44] bought a Mac Mini to show that like any [41:47] any public known is able to connect to [41:48] our server and play play play on our [41:50] server, interact with each other. [41:52] They're like all like working on stuff. [41:54] And we are like putting out like tasks [41:56] that like people would want to like [41:57] these bots incentivize we will be paying [41:59] them out. First person like get all full [42:02] iron armor or like first person to go to [42:04] the nether, something like that. [42:06] And we pay these bots to do that. And so [42:08] each one of these bots is a is a [42:11] Openclaw like sub-agent that is like [42:13] running to like do autonomous like [42:14] actions and you can give them sold MDs [42:16] to like give them like different [42:17] personalities of like what they want to [42:19] do and characteristics. Yeah, we're [42:20] basically building like a whole system [42:23] that's able to help you sleep better. [42:24] So, it has like a different sensors such [42:27] as like a muscle sensor for like [42:30] tension sensor for like detecting if [42:32] you're clenching your teeth, which is a [42:33] big problem right now. Well, especially [42:36] in engineers. We're always stressed. So, [42:37] we're always clenching our teeth. So, we [42:39] have heart rate as another multimodal [42:41] input. For that we use like Apple Watch [42:43] data. We are sending the data from Apple [42:45] Watch to our server and using that heart [42:48] rate along with the EMG signal that we [42:51] are getting from a Myo sensor. [42:53] We try to create an analytics dashboard. [42:55] So, what it does is like first is the [42:57] data collection step. So, once we have [42:59] the data collected using the collected [43:01] data, we actually generate the analysis [43:04] and once we do that analysis, you can [43:06] actually have a conversation with our [43:07] chatbot to actually understand what that [43:09] report is about. So, that it makes it [43:11] easy for the user to understand and the [43:13] third step would be basically [43:16] closing the loop. [43:17] >> Tuna and I suffer from [43:18] like teeth clenching a lot. So, [43:21] basically we have the sensor in like a [43:23] specific muscle and whoop. Whoa. [43:29] Falling. [43:30] >> [laughter] [43:32] >> That happens. Yeah, robots, right? [43:34] >> Yeah. [43:35] >> [laughter] [43:35] >> I think he's aggressive. That's funny. [43:39] All right. Sorry, you're up. [43:40] >> Yeah. So, here our Tuna our [43:43] >> [laughter] [43:44] >> So, basically when he clenches the [43:46] teeth, you can see like how it starts [43:48] like rising. So, you can see like [43:49] there's a muscle like tension. Yeah. And [43:52] that's basically being sent to the like [43:54] our website. [43:56] You'll find like real time analysis [43:58] that's being done over here and once you [44:00] have the real time analysis, we can [44:01] actually generate a report and we have a [44:03] AI chatbot that's sitting on the corner. [44:06] You can interact with it. [44:08] So, you can ask questions regarding your [44:09] report and also book appointments. And [44:11] that's like a closed loop that we were [44:13] talking of. So, my name is Joshua. I go [44:15] to Stanford. [44:16] >> Yeah, okay, cool. So, tell me a little [44:17] bit about what you built here. So, what [44:19] we're building is basically a Zoom AI [44:21] companion that helps you learn and gives [44:24] you animated videos to help you learn as [44:26] well. [44:27] Also, customize practice problems based [44:29] on whatever your lecturer or teachers [44:31] teaching you. And this is just one of [44:33] our features. Uh this isn't our actual [44:35] Zoom AI companion. And so, basically, [44:37] you will collect context from the actual [44:40] Zoom call itself and it'll automatically [44:42] make generate you like a math or physics [44:45] or linear algebra or probability based [44:47] video [44:48] and give it to you in in addition to [44:50] what you're already learning to help you [44:51] learn. And so, basically, any context [44:54] that we collect from the Zoom call will [44:55] be sent to an API and will generate the [44:57] animation. And so, after a while, [45:00] uh we'll get a video like this [45:03] created completely by Wow, looks like [45:05] very much like the three brown one [45:08] brown. [45:08] >> Yeah. [45:09] >> Three blue one brown. He's got like the [45:11] open source kit that you can use for [45:13] making videos. Yeah, and so, we able to [45:15] use AI to generate these videos using [45:17] that package. So, you would see this in [45:19] the side as your teacher's teaching. [45:20] Just another way to help you learn, pick [45:22] up concepts really quickly. [45:24] All right. So, we got 12 hours to go. [45:26] How's the project coming along? What are [45:28] you building? Yeah, the project's going [45:30] pretty good. We are building the [45:31] prediction market for TreeHacks [45:33] specifically. Uh people can go on and [45:35] bet on the events happening. And [45:38] eventually, when the devpost is going to [45:39] be released that people can bet on the a [45:42] grand prize who's going to win the [45:43] actual prize for the TreeHacks. Yeah, [45:46] it's like the Hayekian idea. The [45:48] Hayekian idea is like people [45:50] individually, they don't know that much [45:52] about the world. But together as a [45:54] crowd, [45:55] by voting, by putting money where their [45:57] mouth is, they can contribute [45:58] information to the markets. And as a [46:00] whole, we have the wisdom of the crowds. [46:02] Why did you decide to build this kind of [46:04] an event-specific prediction market? [46:06] >> Right. Well, firstly, we don't think [46:08] that traditional prediction markets work [46:10] for this like fine-grained like [46:12] event-driven thing because there are [46:14] simply so many things happening on the [46:15] ground. And like prediction markets are [46:17] broad, they're for like for general [46:18] audience. So, they can't do this. The [46:21] reason why we want to to wanted to build [46:22] this is because we think that hackathon [46:24] projects, they're usually quite [46:25] ephemeral. [46:26] They're almost like isolating in a way. [46:28] You lock yourself in a room for 36 [46:30] hours. And so, hopefully we can bring [46:31] people together with the fun of [46:33] prediction and trying to contribute [46:35] information to the markets and make them [46:37] more efficient. And we have the team who [46:39] has secured the nicest space. How did [46:41] you get this space? You got an entire [46:42] freaking office room. It was me. So, we [46:44] showed up really early. [46:46] And then we just jumped in the check-in [46:47] line as soon as we possibly could. And [46:49] got out of that as soon as we possibly [46:50] could. And then just ran over and found [46:52] the best room that we could immediately. [46:54] Well, what have you all been working on? [46:56] So, essentially, you know how [46:58] private schooling or like getting a [47:00] tutor is really expensive? It's [47:02] especially in like North North America, [47:03] right? So, essentially what we build [47:05] built was uh is a tool a tutor AI tutor [47:08] that [47:10] uses the Zoom SDK. And you can actually [47:11] like video call it on our app. And it [47:15] will just essentially [47:17] it will kind of like help you learn [47:19] whatever you want to. And I think the [47:23] best way is just to show you guys how it [47:24] works. [47:24] >> us. All right. So, I'm going to join the [47:26] session. And hopefully it works. [47:28] Hey there. I'm Minerva. And I'm really [47:30] excited to help you learn something cool [47:32] today. What's on your mind? Is there a [47:34] subject or topic you're curious about? [47:37] Yeah, can you explain the [47:39] Pythagoras theorem for me? Sure. So, the [47:42] Pythagorean theorem is about right [47:44] triangles. You know, triangles with one [47:45] 90-degree angle. Have you ever seen one [47:47] before? No, I have not. Can you show me? [47:51] Perfect. Let me draw one for you right [47:53] now. [47:54] So, yeah, it can So, it has like full [47:56] access to like whiteboards, [47:58] a graphing calculator, and for example, [48:00] like a 3D renderer. Like, can you show [48:02] me the solar system? [48:08] And it takes a second, but [48:12] wait for it. [48:13] There we go. Let me switch gears and [48:15] show you the solar system. That's [48:16] awesome. [48:17] >> And this is just all generated on the [48:19] fly. Wow. All right, T T minus 12 hours [48:23] to uh submission time. [48:25] What have you all been building, and [48:27] what do you have left? Yeah. Um our [48:29] project is called Collab. It's an AI [48:31] chemistry lab that brings agentic [48:33] intelligence into the laboratory space. [48:37] Uh we were inspired by thinking about [48:39] how great agents have become and where [48:41] we can now deploy them. Our project [48:43] essentially takes standard chemistry [48:45] equipment and standard electronics, [48:47] bringing these two domains together, [48:49] allowing us to automate lab work and [48:51] help researchers focus on what's [48:53] important, right? Which is high-level [48:54] experiment direction, finding novel [48:56] ideas, not pipetting, not working with, [48:59] you know, you know, cell cultures. Yeah, [49:01] of course. If you take a look at this [49:02] table, this is what I've been working on [49:04] throughout these hours. Right now, we [49:06] have a whole hardware setup for [49:09] our experiment that we're going to run, [49:10] and we have Raspberry Pi [49:12] that controls everything here. And I can [49:14] show you a quick demo of of everything [49:16] that's happening here. We have a thermal [49:17] camera that is able to detect the [49:20] temperature when the experiment runs. We [49:21] also have an optical camera, which will [49:23] use to run computer vision. [49:25] Uh on top of that, we have an arm here [49:27] that controls, [49:29] um you know, taking flasks off, putting [49:31] flasks on. [49:32] Um that way, you know, researchers don't [49:34] have to get their hands dirty um doing [49:36] everything [49:37] uh menial in the experiment. So, what [49:39] we're trying to do here is to use AI to [49:41] show [49:42] to to generate board games. So, [49:44] basically, we're thinking of like, if [49:46] you have ever played Monopoly Deal, but [49:47] you want to change like something about [49:49] it. For example, play Monopoly, but I [49:51] want a market crash at the end. So, [49:53] something is a new creative twist on it [49:54] on on games. So, what if you can use AI [49:57] to basically generate these games so [49:59] that you can play it online with your [50:00] friends. So, actually our team here is [50:02] all quite international. We all come [50:03] from different countries. So, like if [50:05] you are able to digitize board games, [50:06] it's also a social element as well cuz [50:09] then we can just continue playing it [50:10] after the hackathon. Right now, we we [50:12] have like generalized the the AI to be [50:14] able to work for a few different games [50:15] like for Uno, Exploding Kittens, and [50:17] poker. So, if like you can just put a [50:19] name, and it creates a room. And then [50:22] from this room, different people can [50:23] like um you can join and you you create [50:25] a board game. So, you can see all the [50:26] different cards for both different [50:28] players. And so, all of this is [50:30] dynamically generated. It's actually um [50:33] based on like using AI to like search [50:34] about the context of the games. Yeah, we [50:36] got 12 hours left. What are the main [50:38] things you're thinking of at this point? [50:39] What are you all working on? I think we [50:40] are like 50% done. So, I think it's [50:42] really a huge mad rush. I think the past [50:44] like 24 hours were mostly like [50:46] experimentation, learning new things, um [50:49] talking to sponsors, and learning off [50:50] their different like technologies. So, [50:52] yeah, right now it's what we're trying [50:53] to do is just finish everything up, [50:54] polish it up, and like get ready for the [50:56] pitch. My partner, Siri, she could tell [50:59] you that she was definitely freaking out [51:02] a bit in the morning, but I told her to [51:05] lock in, and you know what? We did. [51:06] Okay. So, you were worried, and you just [51:09] said, "Lock in," and everything worked [51:11] out great since then. Well, I think like [51:13] I'm the type of person that like gets [51:15] super stressed out over like a the [51:16] prospect of like a huge idea like when [51:19] there's so many different working parts, [51:20] and like I think Jennifer is the type of [51:22] person to be like, "Okay, like chill out [51:24] and like take a deep breath and just [51:26] start working." So, I think that's [51:28] definitely helped. So, tell me a little [51:30] about what you've been building. Yeah. [51:31] Okay. So, for some context, like I run a [51:35] clinical research study on POTS [51:37] patients, but I come from a computer [51:39] science background. So, getting over [51:41] that barrier of like [51:44] understanding all these medical terms [51:46] and things like that has been super [51:47] tough for me. So, like Jennifer, do you [51:49] want to kind of explain more about [51:51] >> Yeah, to speak more broadly about the [51:52] general audience, [51:54] I we both feel like there's a lot of [51:57] people who are really skilled in tech [51:59] and AI these days and who are also [52:02] interested in the medical field, which [52:04] is like how Siri started out. Like she [52:06] just discovered pots and wanted to do [52:08] research on it. But it's really hard to [52:11] be caught up to speed with the medical [52:13] field due to how [52:15] deep and intensive the knowledge base is [52:18] tend to be. So, our goal is to allow [52:22] students or like people like us who um [52:26] to explore medical subfields and the [52:29] most recent clinical developments [52:32] um without like [52:34] as frictionlessly as possible. [52:37] And we're doing that through short-form [52:40] content, which you can elaborate on. We [52:42] basically like uh pull from a bunch of [52:45] clinical research studies that have been [52:46] done in the past and we summarize them [52:48] in a way like kind of like like we we [52:51] want to say like Tik Tok like reels as [52:53] where anyone can go through and scroll [52:55] and like see like okay, what is [52:56] happening in this research paper and it [52:58] like breaks down exactly the complicated [53:00] terms because think about it like in 5 [53:02] 10 years, we're going to have a whole [53:04] group of students who know how to use AI [53:06] and who know how to use these AI tools, [53:08] but the actual like getting over the [53:09] barrier of like using those skills for [53:11] medicine and like combining those fields [53:13] to make some actual progress is like [53:15] kind of what we're targeting. Like we're [53:16] trying to show that it's um it's not [53:18] going to be like these terms aren't [53:19] going to be that scary and there's so [53:20] much like I just feel like there's so [53:22] much clinical data out there that like [53:24] we don't hear about and it would be [53:25] interesting to see like who who there [53:28] with the AI skills can actually do [53:29] something with that. Yeah. And for our [53:32] front end, [53:34] we have our uh [53:37] our model synapsing the gap between AI [53:40] and health care. So, we start off with [53:44] this home page in which the user has um [53:49] two [53:51] question fields. They can answer what [53:53] their interests in the medical area is [53:56] and then what their strengths or like [53:59] interests in AI are as well. So, I might [54:02] put something down like medical imaging [54:07] and then computer vision. [54:10] And then [54:12] by clicking on discover research [54:14] this is where we'll query our back end, [54:17] our server, and [54:20] collect a bunch of relevant articles and [54:24] those relevant articles will then be [54:28] um synthesized into these short-form [54:32] videos that the user can scroll through [54:34] and quickly [54:37] access summaries that are like [54:41] tailored towards their knowledge base [54:43] and experience. On the back end side, so [54:45] we do a bunch of stuff in the back end. [54:47] Like we connect it to like basically [54:49] this vector database we can get the [54:51] relevant information and then we connect [54:53] that to Perplexity and then we connect [54:54] it to something called HeyGen and you [54:56] can kind of see like we we generated a [54:58] video here. So, like it like puts like [55:00] the information on the screen and then [55:03] it makes like the diagrams and kind of [55:05] like in a more explainable way. Awesome. [55:07] Well, keep the great work. Uh what are [55:10] what are your biggest concerns for the [55:11] next 12 hours? I think staying awake [55:13] hopefully and [55:15] yeah, I think staying awake [55:17] and getting this done cuz we're we're [55:18] going to pull an all-nighter. It's the [55:21] it's the TreeHacks way to pull an [55:22] all-nighter. A lot of people waste too [55:24] much time [55:25] uh on dating. Uh I cannot say for myself [55:27] because I have a beautiful girlfriend, [55:29] but uh [55:30] a lot of people spend too much time [55:31] dating, you know, getting to know people [55:33] and uh that's a very time-consuming and [55:35] energy-consuming activity. So, what we [55:36] do is we create a clone of you. Uh so, [55:39] your AI agent can talk to other AI [55:41] agents and get to so, they get to know [55:43] each other and you don't have to waste [55:45] time, you know, trying trying over and [55:46] over again. And you know exactly who's [55:48] like a better match for you to be your [55:50] friend, to be your uh partner, or [55:52] whatever. I think that's exactly the [55:54] subject of like a Black Mirror. Yeah, [55:56] we've got we've got that very [55:57] frequently, yeah. But we we did it in a [56:00] while, you know, more bright colors, so [56:02] it doesn't, you know, give off bad [56:04] vibes, you know? How do you get the [56:06] clone of the person? Like do you have it [56:08] interview you and then turn around and [56:09] pretend to be you, basically? So, it's a [56:11] very very friendly, you know, sign up [56:13] page. Just Currently, we have LinkedIn, [56:15] but uh soon you'll be able to put your [56:17] you know, any link you want. It would be [56:19] your Instagram, your Twitter, and uh [56:21] it'll get the information for you and [56:23] you just write up, you know, a quick [56:24] blurb about what do you like like what [56:26] do you like to do. And uh that's how we [56:28] create your clone and then it instantly [56:30] starts interacting with other, you know, [56:32] agents. What are your main goals for the [56:35] next 12 hours? So, the back end is [56:37] working uh perfectly. The thing is uh we [56:39] broke some stuff in the UI, but after [56:40] that we we have some cool interactions, [56:43] you know, like the agents talking to [56:44] each other. They have actions they can [56:46] do like on a physical not real physical [56:48] world, but they can, you know, shake [56:49] hands, give hugs to each other. Uh [56:51] and then uh an orchestrator gives a [56:53] final decision like should you spend [56:55] your time meeting this person or talking [56:57] to this person IRL? And that's like the [56:59] wow moment that we're looking for. [57:06] All right, you all are 35 hours into [57:08] this 36-hour hackathon, ready to submit [57:11] to judging, and you have a working [57:13] prototype or working essentially what [57:15] you're going to submit, ready, right? [57:17] Let's talk about it. Yeah, so we're [57:18] making a mind assist. It's a mind [57:20] control mind control robotic arm, so we [57:22] can have people who have disability we [57:24] have a hard time feeding themselves a [57:26] hard time [57:28] eating themselves without assistance. [57:29] So, we want to to it [57:30] basically we made a system arm that you [57:33] can use your mind to basically control [57:34] it and [57:36] help feed yourself with the without any [57:38] assistant. So, here's Jessica. We can do [57:40] a demo. Basically, we have two [57:42] components. We have the EEG, so [57:44] basically you put it on the [57:45] on your cortex and it reads a brain [57:47] electrical signals and we basically [57:49] decode it to control the robotic arm. [57:52] And this is the robotic arm and [57:53] basically we have a computer vision as [57:54] well. So, if you see on the left top [57:56] here, we have a computer vision where it [57:58] process [57:59] object detection and it tries to grab [58:02] the food object and then pick it up. And [58:04] so And this is the um [58:07] uh EEG signal. So, basically we have the [58:09] onto waves here and we have to basically [58:11] process them individually. [58:13] Um I mean, it's basically relative value [58:15] to each other and basically we process [58:17] them to this clean data here and we use [58:19] this to basically identify whether [58:21] you're focusing or not. And we use that [58:23] to basically um control the robotic arm. [58:26] So, we made a pair of smart glasses that [58:29] can help you buy anything you see and [58:31] anything you want. Hey, I'm Bella. [58:35] Hi there. How can I help you today? Hey [58:37] Bella, can I buy this bag of granola? [58:41] Sure. Let me look at that and find the [58:43] best matches for this granola. Give me a [58:45] moment. [58:47] Okay, so currently the smart glasses are [58:50] it's capturing an image and it's using [58:52] computer vision to essentially do like a [58:53] look up and try to find like a [58:55] product. And currently it's heading our [58:57] back end. This is Dot and what Dot aims [59:00] to do is bridge the gap between elderly [59:04] people and the modern ecosystem of how [59:07] medicine works. And so, that means that [59:10] patient charts are built for, you know, [59:13] mobile apps and all of these things that [59:16] don't really work very well for people [59:18] who are on on the older side. And so, [59:22] for us [59:23] that looks like [59:25] building a voice assistant and taking [59:27] all of those interactions that aren't [59:29] designed for them and reframing those [59:32] into things that they can sort of [59:34] natively engage with. [59:35] >> Okay, cool. Yeah. So, how does it work? [59:38] This is uh Dot. You can see inside here [59:41] we have speakers, microphones, and [59:44] various other uh things. You can see in [59:47] the top here, there's a scanner and that [59:50] scans barcodes on medicines. And so, [59:53] there's a lot of small text on here that [59:55] is really hard for people who might, you [59:57] know, be visually impaired to to read. [60:00] But, medicines all have barcodes on them [60:02] and those either correspond to universal [60:04] drug codes, universal product codes, or [60:08] to the prescription numbers. And so, we [60:11] have uh AI agents that run as a cron job [60:14] and will scrape their patient portals [60:17] for information on their prescriptions [60:19] and medical history. [60:21] And so, [60:22] uh [60:22] stored securely locally on device. [60:25] There's always basically uh a cache of [60:29] their patient data. [60:31] >> Okay. Well, why don't you scan and and [60:33] show us what the core interaction is? [60:35] If we um [60:37] scan something here, we can say, "Hey [60:40] Jarvis, tell me about my heart [60:42] medications." [60:43] Leo, your heart medications include [60:46] metoprolol, [60:47] atorvastatin, and aspirin. [60:50] They help manage your heart rate, [60:51] cholesterol, and reduce clotting risk. [60:55] Perfect code camp? Yeah. Hell yeah. So, [60:57] the problem is that coding right now [61:00] um using agentic coding tools is very [61:02] supervised. Um you have to go back and [61:04] forth with Claude, back and forth with [61:05] Open Claude. Some people do this [61:07] unsupervised, they let their agents run, [61:09] but for longer running tasks, this [61:11] breaks and the quality degrades over [61:13] time. So, essentially just more cohesion [61:16] uh when [61:17] >> for extremely [61:18] We believe in the future in like a [61:19] couple months, like you'd be able to [61:21] build a browser in one shot. You'll be [61:22] able to build a compiler in one shot. [61:24] And we're we're doing some research into [61:26] like long-running agents. This is like a [61:28] UI that you can visualize what would [61:30] like the idea happen behind it. So, [61:32] essentially the goal is that you just [61:34] don't need to supervise nearly as much. [61:36] Micromanage. So, we spin up a swarm of [61:38] agents, about over 100 coding agents, [61:41] and these commit at over 1,000 commits [61:43] per hour. And so, our previous proof of [61:45] concept was to build Minecraft in one [61:47] prompt. And so, this is doing a replay [61:49] of that run. 3 2 [61:51] 1 So, you are on 2 hours of sleep. [61:54] >> I am. You've been grinding, getting [61:56] things done. Tell Tell me a little bit [61:58] about what you built here. So, we built [61:59] Recite. Um essentially, we want to [62:01] bridge the gap between uh browser [62:03] accessibility for individuals with [62:05] visual impairments. Um basically, like [62:08] you can imagine Facebook and like [62:09] Instagram has a lot of like very cool UI [62:11] features, um like short-form content, [62:14] but a lot of that is inaccessible to [62:16] people that are like blind or have other [62:18] visual impairments. And so, we want to [62:19] make a browser or we want to make an app [62:21] that makes it more accessible for people [62:24] that are blind to use these browsers and [62:25] things like that. For instance, I'm [62:27] going to close my eyes and I'll just [62:28] talk to you. I'll say, [62:29] "Hey, um could you show me the [62:31] differences between Stanford's [62:32] architecture with San Jose State's [62:34] architecture? Um I'm particularly [62:36] interested in seeing which one has older [62:39] architecture." I'll click this and it [62:41] sort of starts processing. And so, [62:43] there's a bunch of AI agents that we [62:44] have. We have diagrams that visualize [62:46] what's going on in the background. But [62:47] technically, obviously, if you're [62:48] visually impaired, you won't actually [62:49] see this in the background. This is for [62:50] just judging purposes, right? And so, [62:52] what essentially happens is our app sort [62:54] of just routes instructions to an [62:56] underlying browser. The browser then [62:58] performs the searches for us to answer [63:00] ultimately the question that we just [63:01] asked. [63:01] >> Okay, interesting. So, you send it off [63:03] to go ahead and essentially research [63:05] Yeah, exactly. [63:06] >> like a human would. Now, we have an [63:07] answer. [63:08] All right, cool. Founded 1857 versus [63:11] Stanford's 1891. But Stanford's more [63:14] famous for that iconic Mission Revival [63:16] main quad with red tile roofs. SDJ makes [63:18] it Kevin Gu, how's your weekend been so [63:20] far? Yeah, it's been great. Um crazy as [63:23] awesomely hackathon. Like everything's [63:25] been really fun and been welcomed since [63:27] day one since we flew from Canada. So, [63:29] it's been awesome. Great. [63:30] >> Well, glad to have you back in safe [63:32] view. We're of course at CalHacks as [63:34] well. And tell me a little bit about [63:36] what you built here. Yeah, definitely. [63:38] So, for us we wanted to build something [63:40] for the elderly. We felt like something [63:42] that has happened is that as time has [63:44] gone to more complex tech has evolved, [63:46] but unfortunately that meant that the [63:47] seniors got left behind. [63:49] I've been there to help my grandma and [63:50] my parents-in-law with their tech, you [63:52] know. I once actually drove 30 minutes [63:54] to go visit my grandma to take 2 minutes [63:56] to set up her Wi-Fi, drive back home. [63:58] So, for me kind of with that we wanted [64:01] to improve that. And we also noticed a [64:03] real problem now, which is that actually [64:06] there's going to be a big aging [64:08] population crisis. One in six people [64:10] actually by 2030 will be over 65 and [64:12] that'll double by 2050. So, for a lot of [64:14] health care suddenly we want to urgently [64:16] provide them the support they need and [64:17] innovate on how we can support health [64:19] care there. So, our solution we built [64:21] essentially a conversational AI [64:23] caretaker that's as easy to talk to a [64:25] friend. So, you don't need to go through [64:26] some sort of complicated app or download [64:28] apps for the elderly. You can just [64:30] simply talk to it. There's going to be [64:31] two sides of it. First is going to be [64:32] kind of a dashboard for the caretaker to [64:34] see. But second, we also have a really [64:36] really simple interface for the elderly [64:38] to see as well. So, I'll just talk to [64:40] it. [64:41] Very very simple. [64:43] I'll say, "Hello. It's uh good morning. [64:45] I hope you're doing well and uh [64:48] how's how's you have you been?" [64:50] And it's going to be an AI that actually [64:51] talks to us. And through a lot of [64:53] research we actually learned that [64:54] elderly people actually do prefer [64:55] talking with AI. And unfortunately, you [64:58] know, sometimes they may feel lonely. [64:59] They unfortunately might be widowed. So, [65:01] for them talking to an AI for them is [65:03] actually better than talking to nobody. [65:05] Um and for us we actually have something [65:08] running in the background where we're [65:09] analyzing this call transcript it time [65:11] to detect symptoms of cognitive decline. [65:14] So, for those who may be kind of having [65:16] early symptoms of dementia, we're [65:18] analyzing the transcript to really see [65:19] that. [65:21] Hey, how are you all feeling? Very [65:23] awesome. Very tired. You ready to demo? [65:26] Sure. Let's demo. So, basically we [65:28] realized that like a lot of startups or [65:30] smaller companies have a really big [65:31] problem hiring people, especially [65:33] nowadays like applications are so volume [65:35] based. So, like when you open up a [65:36] posting, you might be a small company, [65:37] but you'll see like hundreds and [65:39] hundreds or thousands of applications [65:40] and it's really hard to like go through [65:42] and figure out what applications are [65:43] actually good and what aren't. So, [65:45] basically we built a tool to help that. [65:46] It kind of uses a lot of like multimodal [65:48] models and stuff like that. But, the big [65:50] thing is at the end we have a big AI [65:52] agent which kind of like aggregates and [65:53] filters down based on your specific [65:55] needs. Uh how does it work? Basically, [65:57] once you do a posting, you you are like [66:00] assured that your posting will only go [66:02] out to people who are actually qualified [66:04] and interested and also good fits. And [66:06] after this, after you get a whole bunch [66:07] of applications, you can click the score [66:08] now button and this sort of just like [66:10] ranks intelligently ranks and sort of [66:12] like breaks down all the candidates so [66:14] you can be confident that you get the [66:15] right people for the interviews. So, [66:18] like for example, you can look at this, [66:19] it gives you a score, you can look at [66:20] the resume, and then it also gives you [66:22] like a more intelligent breakdown on [66:24] like skill set. [66:26] >> So, a big thing for us is we focus on [66:28] showcasing only jobs that we think the [66:30] candidates are fit for. And on top of [66:32] this, we also sort of limit the number [66:33] of applications. So, if if you can [66:35] imagine in a perfect world, candidates [66:37] will only apply for a handful of [66:38] companies that they're good fit for and [66:40] they know and then off of this, everyone [66:42] will have the same amount of jobs going [66:43] around, right? But, the process is just [66:45] simplified a lot more. So, that's sort [66:46] of the goal here. B. Oh, okay. [66:49] Middle one. [66:57] My understanding is this [66:59] gives you chord suggestions. [67:01] >> Yes. Depending on what you played [67:03] previously, it'll [67:04] it'll give you the next chord and you [67:05] can just play that [67:06] >> Yes. Uh based on that. Like kind of like [67:07] almost like sheet music that is being [67:09] dynamically written as you go. [67:11] >> Yes. Yes. Yes. How these came along. Why [67:13] did we decide on these ones? So, our [67:16] approach is that you have all the [67:18] possible combinations of the chords in a [67:20] certain octave. So, we treat notes that [67:23] are on different octaves [67:25] as the same. These are called chroma [67:26] vectors. It's a [67:28] array of 12. So, five black keys, Right. [67:30] >> seven white keys. And then so you have [67:32] all the possibilities in that and we [67:33] store it into an array. And then what we [67:36] do is we process different signals. We [67:38] look at all the different intervals that [67:39] pass through this chord. You could have [67:41] like a third, the fifth, where you have [67:43] a tritone and we see how much of each [67:45] one passes through the signal. And this [67:47] is just done with a Fourier transform. [67:49] And then we put that into a [67:50] high-dimensional space. This one is in [67:52] 24 dimensions because you have 12 types [67:54] of signals, each one in a complex [67:56] number. So, a complex number can be [67:57] represented with two real numbers. [67:59] That's going to be 24 dimensions. And [68:00] this space, geometrically, it preserves [68:02] some musically like interesting things [68:04] like what key you're in. Or the angle [68:06] between certain vectors can can uh [68:08] describe how consonant they are. And [68:10] then using this, [68:11] the chord suggestions that we produce [68:14] will describe like a certain amount of [68:16] tension depending on where they are in [68:17] the space. So, then you would traverse [68:19] through the space adding tension, [68:20] releasing tension, adding tension, and [68:22] releasing tension. This is kind of our [68:23] prototype to that. Very cool. Yes. All [68:26] right. Yeah. So, our project is called [68:28] Frontline and essentially it's [68:29] revolutionizing the idea of how [68:31] operators respond to like emergency [68:34] calls, for example. Um it takes around [68:36] like 7 minutes for a first responder to [68:38] show on site and there are a lot of [68:40] things that occur in those 7 minutes [68:42] like getting the information, getting [68:44] details about the scene. Like if you see [68:46] a car crash, for example, you're [68:47] reporting a lot of things, but we have a [68:49] lot of technology and tools to help [68:51] assist with that, provide our first [68:53] responders some details before they come [68:54] onto the scene, right? So, we've kind of [68:56] built this little pipeline that you can [68:58] basically phone call an operator using [69:01] like Zoom, but it's just a video call [69:03] and actually it can scan your [69:04] surroundings and analyze it for you. So [69:06] if you have an image of a car crash or [69:08] like there's a car crash, [69:09] you know, if I'm calling the operator [69:11] being like, "Hey, I see a car crash. [69:13] There are two cars completely totaled." [69:15] It'd be able to record my speech and [69:18] then be able to note down what it sees [69:20] also in the camera and generate a report [69:22] for the operator to send to first [69:23] responders. [69:24] And at the end, we also have an ability [69:28] to get text updates like first [69:29] responders, people, operators can like [69:31] send a message to this like agent to [69:34] provide contextual details like over [69:36] like messaging about up color upcoming [69:39] instances about the scene for example. [69:41] Yeah, so what was your experience [69:43] overall building all of this [69:45] uh during the hackathon? Like did you [69:47] all achieve most of the functionality [69:49] you hoped to build out? Yeah, I think [69:52] that we had our certain things of like, [69:54] "This is what we need." Like our MVP of [69:56] like what we need to make our thing [69:58] succeed. This is what would be cool to [69:59] have and this is like kind of our like a [70:02] vision. And then I think we kind of [70:04] reached the cool to have area. I think [70:06] like there's parts we definitely could [70:07] have improved on, but then we achieved [70:10] all of our basic things of like complete [70:12] analyzing like the imagery, being able [70:15] to get a transcript and generate a [70:16] report using AI and being able for two [70:19] people to simultaneously do that and [70:20] call each other at the same time. [70:22] So what are we looking at? Yeah, so um [70:24] one of the new upcoming fields that's [70:26] happening right now is this uh you can [70:28] see war that's happening on the ground, [70:30] in the water, in the air, but soon space [70:31] is going to become a pretty contested [70:33] domain. We have a lot of spacecraft that [70:34] are in orbit and we're able to track [70:36] these spacecraft on the ground and if [70:37] there's ever a case where the spacecraft [70:39] might crash, we can send those signals [70:41] up to the spacecraft to move them. But [70:42] what if there's no longer you can no [70:44] longer talk to a spacecraft? So there's [70:46] already examples where there's there's [70:48] jamming active jamming happening where [70:49] countries are actually jamming signals [70:51] from for spacecraft and you can no [70:53] longer talk to them. And so to do this, [70:54] we basically have to find a way for the [70:56] spacecraft to have her own space traffic [70:58] management system and track how all the [71:00] satellites move around. So, this is a [71:02] very complicated problem, but by [71:03] leveraging NVIDIA GPUs, we're actually [71:05] able to run this on the spacecraft [71:07] itself, propagate thousands of [71:09] satellites, and figure out if a [71:10] collision will occur. So, here we can [71:12] actually see these are star tracker data [71:13] we're simulating. These are basically [71:14] cameras on satellites that actually look [71:16] out and look for individual like dots in [71:20] dots in the sky like stars, but we're [71:21] repurposing them to actually track the [71:22] satellites. And so, by tracking these [71:24] satellites, we're able to actually pick [71:25] out how they move, and we can propagate [71:27] this in real time, and we can actually [71:28] predict if a collision would happen. So, [71:30] here you can actually see we found a [71:31] collision. [71:32] >> A collision of two satellites in space. [71:33] >> Exactly. So, if there's a piece of space [71:35] debris [71:35] >> retro thrusters or something they can [71:38] like adjust their [71:40] orbit slightly Exactly. So, we have like [71:42] a case where they have rocket engines on [71:43] the spacecraft itself. You can actually [71:44] if you can predict a collision will [71:45] happen, you can then [71:47] apply evasive maneuvers to get out of [71:49] the way. But, one interesting [71:51] interesting thing that might come up is [71:52] is a chaser and evader scenario. So, [71:54] let's say I'm a Taiwanese uh [71:56] telecommunications satellite, and China [71:58] wants to come and sabotage us. So, they [72:00] might be coming to collide with us or [72:02] basically rendezvous and try to disable [72:04] us using like a claw or some other other [72:06] like weapon. And so, what we can do is [72:09] if we can see they're coming at us, we [72:10] can actually try to move around. And so, [72:11] here you actually have a chaser evader [72:13] scenario where we basically [72:16] uh try to avoid the satellite, and then [72:18] the satellite is following us, and we [72:19] have to keep moving around until either [72:21] they run out of fuel or we're toast. [72:23] Wow. So, what did you build over the [72:25] course of the weekend? Yeah, so over the [72:26] weekend we basically built this orbital [72:28] propagator that we're running on a GPU. [72:30] We built this uh star tracker [72:31] measurement system that can actually [72:32] track the satellites themselves, and [72:34] then we've built a prediction algorithm [72:36] that can forward predict and solve for [72:37] that. We've also implemented a common [72:39] filter [72:40] a a common filter for for the [72:41] estimation, and also a convex [72:43] optimization program that's actually [72:44] going to then go and try to solve for [72:46] the next best orbit to actively evade [72:48] and make sure we don't have any [72:48] collisions. I've overall I've interested [72:50] in machine learning. I mean, I taught [72:52] myself to code at a really young age [72:54] using Scratch. So, I thought why not [72:56] make something that helps out kids these [72:57] days? And, you know, the next big thing [72:59] after computer science is machine [73:00] learning. So, all of us decided to make [73:02] Scratch for machine learning, [73:03] essentially. And we made Burn. What you [73:04] can do on this website is you can go [73:06] through a bunch of different types of uh [73:08] lessons and presets that allow you to [73:10] learn and understand exactly how [73:12] different machine learning algorithms [73:13] work. So, you got neural networks on [73:15] here, random forest models, um you also [73:18] have large language models, uh [73:19] visualizations for things like Chat [73:21] GPT-2. [73:22] Um [73:23] and most notably, the one that we are [73:25] really proud of is AlexNet. Uh so, if [73:27] you guys have heard of AlexNet, it [73:29] allows you to use a CNN to classify [73:30] images. And this one, what it does is it [73:32] classifies images of dogs and cats. So, [73:34] here's our um [73:36] convolutional neural network. You can [73:38] move it around, you can see how things [73:39] interact in space. You can see uh this [73:41] is our input layer over here. [73:43] Um and then you can see all the [73:44] different convoluting layers. And then [73:46] you also have your output layer that is [73:48] just two classes at the end. We'll [73:49] either tell you either it's a dog or a [73:51] cat when you put in an image. Uh and [73:53] then yeah, the entire process of what [73:54] Burn does is it allows you to visualize [73:56] the whole training process. It allows [73:58] you to uh choose whatever parameters you [74:00] want to do for your training process, [74:02] such as your batch size, the number of [74:03] epochs, uh what loss function you want [74:05] to use, as well as what optimizer you're [74:07] using. And then from there, uh you can [74:09] actually visualize the entire training [74:10] process and see how things evolve in [74:12] time. Our application is called Rewrite, [74:15] and basically what it does is it lets [74:17] users upload a picture um [74:19] of like a scene they want, or they can [74:21] also prompt it. But for this scene, we [74:23] uploaded a picture of uh the moon [74:25] landing. [74:26] So, a user is able to upload that [74:28] picture right there, uh describe an [74:30] action, so how they want the camera to [74:31] move. [74:32] And then right now, what we're doing is [74:34] we're generating like a full 3D scene [74:35] that the user can interact with. Um it's [74:38] a little bit limited right now in that [74:40] it's still a 3D scene, but the user [74:42] isn't able to move how they want. They [74:45] They have to describe it ahead of time [74:46] here. [74:48] Um it does take a little bit of time [74:49] since it is a pretty computationally [74:51] heavy task. We could show you a video. [74:53] Yeah, that's great. [74:54] >> Of it already generated. So, just press [74:56] play. [75:02] Yeah, so that would be an example. [75:04] So. [75:05] Wow, and this is all rendered on top of [75:07] that image. [75:08] >> On top of that image. [75:09] >> So, it would it figured out a way to [75:10] kind of like predict what the rest of [75:12] the uh lander would look like. [75:13] >> Exactly, yeah. Yeah, like we're using [75:14] diffusion models to kind of uh expand it [75:17] and we're also generating like point [75:18] clouds. So, you can do like rotations [75:20] and panning to the left or right. Um so, [75:22] it is fully generating a 3D scene from [75:23] the perspective of the user, but only [75:25] like where we're where required. [75:26] >> Yeah. [75:27] Um so, if you want to try the the [75:29] headset, you can kind of feel immersed. [75:31] So, the idea is we'd have like students [75:34] or or I guess anyone who wants to be [75:36] educated being able to relive these [75:38] super cool experiences. [75:40] Very cool. Yeah, I can feel it like [75:42] panning out. Yeah. [75:45] Awesome. Now, well, let's talk a little [75:47] bit about the hackathon journey. What [75:49] has happened over the last 36 hours? [75:50] Obviously, you built all this. Yeah. [75:52] But, what else has happened? We bought [75:54] swords. I won the robot boxing fight, [75:57] which is pretty cool. [75:57] >> the robot boxing fight? [75:58] >> Yeah. This hackathon has really had so [76:00] many like side quests and adventures. [76:02] It's really been one of the coolest ones [76:04] I've been to. [76:05] Um but also, we had like many [76:06] opportunities to really grind together [76:08] for so many hours and very very few to [76:11] sleep. I want to outdoors is that to [76:13] leverage computer vision [76:15] and our visual visual learning models to [76:18] come up with [76:20] workflows that you can [76:21] use to build AI agents quickly. We did [76:24] it overnight. Um he's going to show you [76:26] how that works. [76:28] When you upload a video, it is going to [76:31] um [76:31] basically give you like um use computer [76:34] vision and vision visual language models [76:36] to really pretty much understand what it [76:38] what was going on in your workflow. We [76:40] use our proprietary algorithms to sort [76:42] of like really figure out what is going [76:44] on in each [76:45] each frame and and and like sequences of [76:48] frame. And then what that creates is [76:49] that it gives you the space to generate [76:52] a workflow from [76:54] the video that was going on over here. [76:56] So that's what we have today. [76:57] >> it it'll look at like your data that you [77:00] give it which could be like a screen [77:02] recording and then it'll kind of reverse [77:03] engineer how to do that [77:05] through this process and you can kind of [77:07] go in and scrutinize it. It's not just a [77:09] black box. You can kind of see how it's [77:11] going about approaching all the [77:12] different aspects of the task. [77:17] All right, we're here with our resident [77:18] cyborg. [77:20] Ready to check in and see [77:22] how the muscle movement of the jaw is [77:25] doing. Hello guys. [77:27] What we are building is called Sleep [77:29] Sense. So it's basically like you might [77:32] have seen here a lot of people like who [77:34] are ambitious and hackers, they don't [77:37] get to sleep as much they want. But on [77:40] top of it, they sleep say 5 hours, 6 [77:43] hours or even 10 hours sometimes and [77:45] wake up tired. Wake up feeling pain in [77:48] jaw. People are grinding their teeth in [77:50] the night which causes like jaw pain and [77:52] also disrupts your sleep cycles. [77:54] >> Yeah. And [77:56] that's why we would like [77:58] what if it could be that like we could [78:00] build something that can help people [78:01] diagnose this earlier and [78:04] offer the data to their doctors so that [78:06] they can track their history, offer them [78:08] right treatment. [78:10] And also they know what [78:12] what is the reason behind it because it [78:14] could also be stress induced, it could [78:16] be genetic, it could be hormonal [78:18] imbalance. [78:19] >> Awesome. So let me start talking. Yeah. [78:22] And you can see that it changes to [78:23] talking actually and [78:25] let me show you how it looks when I [78:27] clench. [78:30] What if we could detect muscle failure [78:32] before it's happening? This is actually [78:33] a big issue, especially for example [78:35] doctors are often experiencing muscle [78:36] fatigue in the operating room. [78:39] And this is contributing to medical [78:40] errors and it's a commonly cited in [78:42] malpractice lawsuits. [78:43] And so, we're hoping to predict a [78:46] person's muscle fatigue and their [78:47] eventual failure well before it happens [78:49] and maybe even before they can feel it. [78:51] And this is applicable not only for [78:52] doctors and surgeons, but for anyone who [78:54] works in repetitive, dangerous, and [78:56] stressful actions like construction [78:58] workers or even athletes. [79:00] And so, what we have here is a set of [79:02] electromyography electrodes attached [79:04] here to Eric's bicep [79:06] which is then processed by our Python [79:07] script. And so, we're going to have Eric [79:09] do a static isometric hold of a of a [79:12] dumbbell [79:13] and hold try to hold this as high as [79:14] possible uh without letting it fall. And [79:16] eventually, his muscles will fatigue. [79:18] And by measuring the current running [79:19] through his bicep, we're going to [79:21] predict exactly when he'll drop it uh [79:23] even if he doesn't know. So, let's give [79:25] this to him. I'll hold this off. And [79:26] Drew, want to you want to count it off? [79:28] 3 2 1 go. [79:31] Okay, so right now, this is going to [79:33] calibrate. So, Eric is [79:35] tensing as hard as he can and we're [79:36] measuring uh the maximum strength of his [79:38] bicep contraction. The key parameters [79:40] that we're pulling here [79:42] is RMS, which is representing the [79:44] amplitude of the contractions of his [79:45] muscle fibers, uh as well as MDF, which [79:48] represents the frequency of how fast [79:50] those muscle fibers are activating. [79:52] And by measuring how that MDF value [79:55] decreases over time, we're able to [79:56] predict uh when he's going to drop this [79:58] weight. So, uh in real time, we have a [80:00] prediction of uh in the future when we [80:03] think that weight is going to drop. [80:04] Obviously, Eric doesn't know and we're [80:05] going to see uh as he holds as hard as [80:07] he can how close we are going to get [80:10] there. [80:11] Uh and the way this is working is we are [80:13] bidding this data in in sets uh as we [80:15] progress in time [80:17] and uh just continually updating our [80:20] slope and our prediction for when we are [80:22] going to reach a critical threshold [80:23] value where we think he's going to drop. [80:27] We've done this so many of these trials. [80:30] Okay, so that's when he gives out and we [80:32] see that was within 10 20 seconds off [80:35] here from our prediction, but you know, [80:37] maybe even 50 seconds before he actually [80:39] failed, our prediction was able to [80:40] stabilize and estimate when he would [80:42] drop it. Very cool. How have you enjoyed [80:44] the event so far? So much. Like I've [80:46] done hackathons before, but nothing like [80:48] TreeHacks. I've done like four before, [80:50] but TreeHacks takes it to a completely [80:52] different level. And I think the best [80:54] thing about it was just talking to so [80:56] many people working in great companies [80:57] and companies like I really want to be [80:59] in, companies which are relatively new [81:01] and doing great stuff. So, I really [81:02] enjoyed that. How did it go, man? So, we [81:04] got most of the things we wanted to get [81:06] done done. [81:07] And as we were like kind of recording [81:09] our final demo video, things broke, of [81:11] course, cuz nothing ever goes right [81:13] during a hackathon, but yeah, I think I [81:15] think we've got a pretty good shot. [81:17] Awesome. [81:18] Are there any big takeaways from this [81:20] weekend? Like any like lessons learned? [81:22] I think definitely get some sleep and [81:25] have fun because [81:26] the last few hackathons I've been to, [81:27] I've been just like sitting in one room [81:29] and just like hacking away for 36 hours, [81:32] but TreeHacks is really the first [81:33] hackathon that I've actually like gone [81:34] out and have fun had fun and I've had so [81:36] much more fun and like I [81:38] this is definitely been the best [81:39] hackathon I've been to all all year. [81:41] We're here with the judges. We've got a [81:43] pretty good stable of judges. They're [81:45] going to go out there, they're going to [81:46] evaluate the projects. Talk about how [81:49] you're going to be evaluating these [81:50] projects. Do you have like some rubric, [81:52] some scoring system? [81:54] Uh we do. We have a few things we're [81:56] looking for, I think. Um you know, we [81:58] got to be really wowed by the projects. [82:00] We got to uh you know, it's got to be [82:02] really technically impressive, you know, [82:05] cuz uh you got like 36 however long [82:08] to you know, make something really [82:09] impressive. [82:10] And if they think really outside the box [82:12] and like have like a real like impact or [82:15] whatever, I think that's [82:16] that's pretty cool. So. I like [82:17] creativity. So, I want to see something [82:20] that's really cool, out of the box, [82:21] never thought of before. [82:23] Um maybe something in like film and [82:25] arts, that would be really cool. I mean, [82:26] I think with AI tools, anybody can make [82:28] stuff now. So, I agree on the creativity [82:30] point. I feel like you should really be [82:31] pushing what's possible to make. Like [82:33] something, I don't know. I I expect to [82:35] have my mind blown. This is arguably one [82:37] of the best hackathons in the world. So, [82:39] I hope to see some of the best hackathon [82:40] projects I've ever seen. So, we shall [82:42] see. Uh I want to see something that [82:43] makes me question why there was a box in [82:44] the first place. Uh I'm just a sucker [82:47] for anything fun. So, if it makes me [82:48] laugh or I think it's joyful, I'm like, [82:50] "That's awesome. That's epic." I can get [82:52] on with that. I think something that's [82:53] really fun that makes me burst out [82:55] laughing or is tell everyone tell all my [82:57] friends about it. [82:59] I'm looking for that. So, Thais, you are [83:01] one of the main [83:03] volunteers who put together this entire [83:05] hackathon. It's been incredible. I've [83:07] had lots of people throughout the day [83:08] tell me they've been to like dozens of [83:10] hackathons. This is the best one they've [83:11] ever been to. [83:12] >> great to hear. [83:12] >> Yeah, like incredible feedback from the [83:14] hackers. [83:14] >> Oh, amazing. Yeah, no, we we've spent [83:16] like the past uh 8 months like putting [83:18] this together for everyone. Uh we're [83:21] just a team of like 27 students doing [83:23] this. So, it's been so much work uh and [83:25] we're so glad it turned out great. [83:27] People are having a great time. Yeah. [83:28] Well, tell me what goes into setting up [83:31] an event like this cuz I think people [83:32] underestimate like the monetary cost, [83:34] the amount of volunteer hours, [83:36] >> Yeah. all the different arrangements you [83:37] have to get with like all these [83:38] different sponsors, [83:40] >> Yeah. uh securing the venue, security, [83:42] like all these other different [83:43] considerations that people who might [83:45] want to create like a student hackathon [83:47] at their school that they may not have [83:48] even thought of. [83:49] >> Yeah, for sure. I mean, you're you named [83:50] a lot of the the main things uh us [83:52] organizers have to deal with. A lot of [83:54] uh our logistics co-director, Rachel, [83:56] like right before the event, I remember [83:58] her to-do list was like literally three [84:00] pages of Google Docs that she needed to [84:01] get done in the last week of like each [84:03] individual line items with like [84:04] different admins she had to talk to, [84:06] different uh like requests she had to [84:08] put in. Yeah, there's so much to do with [84:10] like renting buildings. Um [84:12] we cover hackers' flights to come out [84:14] here. So, uh we have to coordinate all [84:15] of that. And our I mean, we have amazing [84:17] sponsors, of course. And uh recruiting [84:19] all all those sponsors takes a lot of [84:21] work, as well. But, we have like really [84:23] really incredibly dedicated team members [84:25] that each like bring something they're [84:27] super excited or passionate about. I [84:29] mean, these badges were done by one of [84:30] our team members named Torrin. Uh he [84:32] just was really good at making PCBs. So, [84:35] uh he uh we gave him this this project [84:37] idea and he was like, "Yeah, let's do [84:38] it." Uh and ended up putting this [84:40] together. Um [84:42] And yeah, so everyone sort of like is [84:45] it's it doesn't feel like work because [84:47] in the last like few months, it's kind [84:49] of a break from your studying and you [84:50] end up just spending as much time as you [84:52] can uh working on TreeHacks because it's [84:55] fun. Like, he loves to work on the PCBs. [84:57] I love to uh build the systems that make [84:59] this event run. I love to coordinate [85:01] with admin. Uh do all that. [85:02] >> Yeah, how many more years do you have [85:03] here at Stanford? I'm a sophomore, so I [85:06] have two more years. [85:06] >> Are you planning to be involved next [85:07] year? We have an interesting org [85:09] structure actually where only people [85:11] only stay on the team for uh 1 year [85:14] unless they become a director. So, I'm [85:16] out next year. Uh I'm off the team. So, [85:18] it's it's kind of weird. Like, you would [85:19] think we'd maybe want to keep people on [85:21] the team for longer, but it kind of [85:22] brings a new sort of flow and [85:24] >> Yeah, I mean, turnover so it doesn't get [85:25] stagnant. It probably dramatically [85:27] minimizes politics and stuff like that [85:30] if you know you're on your way out. You [85:32] really don't see it come together until [85:33] you walk in the building and just see [85:35] like, "Wow, there's tables here. These [85:36] these block letters. Uh there's these [85:38] signs. This amazing design that we had [85:40] put together." And like, without every [85:42] single person on our team, it wouldn't [85:43] have been possible. [85:44] Yeah, well, shout out to every single [85:46] person who volunteered to help [85:48] >> Everyone. [85:48] >> make TreeHacks Yeah. [85:50] >> 2026 a reality. Every single person had [85:52] like [85:54] was required to make it incredible. So, [85:55] yeah. Welcome to Stanford [85:58] TreeHacks judging. There are a thousand [86:01] developers behind me. Just an ocean of [86:03] people and they all have table space [86:05] where they're going to be demoing their [86:07] project. Let's go check and see what [86:08] they've been building. Um we're building [86:11] um [86:11] a product where we can play we can use [86:15] any sort of everyday object and turn it [86:17] into any type of instrument we want. [86:20] Currently Mark is trying to show how we [86:22] can create a guitar. Uh we use computer [86:24] vision to analyze the strum and notes, [86:28] and then we use uh audio processing to [86:30] understand how we can like understand [86:32] which note is being played and how to [86:34] convert that into a note we want to [86:36] actually understand. And then we connect [86:38] that with some multimodal um models that [86:41] will analyze um that will analyze how we [86:44] how Mark plays to give him personalized [86:46] data-driven insights and improvements on [86:48] how he can put like potentially like [86:50] improve the way he plays. We also have [86:53] uh connected some APIs where the way in [86:55] which he can play [86:57] um we can convert those notes into text [86:59] that uh Chapel will understand and then [87:02] generate an entirely new song based on [87:04] how he plays. Very cool. So yeah, that's [87:06] pretty much uh the entire kind of uh [87:09] project, yeah. Well done. Yeah, so we're [87:11] building basically a food waste [87:13] detection system. So we have an Intel [87:15] camera right now. Basically it checks [87:17] the food and sees how much of it is [87:18] wasted. So we were partnering with uh [87:20] Stanford dining halls. Basically they [87:22] have a lot of waste and so we went to [87:24] the the dining halls as you can see [87:25] here. Um we have a list of all the the [87:28] meals that we went to. Basically you [87:29] just scan the item right here [87:31] like this. [87:34] It It once it turns green [87:40] and then it tells you right here half of [87:41] the apple was was left uneaten. So it [87:43] kind of gives you some some metrics on [87:45] that. You know that as college students [87:46] it's becoming very hard to make friends [87:48] and current social media feels extremely [87:50] performative and really bland. So we [87:52] want to make a way to find new [87:55] connections and new friends in a way [87:56] that feels a lot more organic. So we [87:58] have wavelength. So as you can see, [88:00] users can post images or text that they [88:02] like that's in very chill and [88:05] non-performative way. And so we use our [88:07] cloud AI API to generate things that [88:10] relate to those. So as you see here, [88:12] Austin Marathon, they say running and [88:13] cardio and endurance sports. Then, [88:16] instead of having to generate things to [88:18] think of whenever you want to text [88:19] someone, when you go to chat, it [88:21] automatically generates a prompt for you [88:22] to connect to others. So, I see that me [88:24] and Letou both like outdoor activities. [88:26] So much of the difficulty making new [88:28] friends is finding what you want to say [88:29] to think you're over it. So, we're going [88:31] to do this for you. All right, so we're [88:33] here with a solo hacker. He built [88:35] everything himself over the course of 36 [88:37] hours. Take it away. Yeah. So, this is [88:39] reasonable graphs. What reasonable [88:41] graphs is is a very specific framework [88:44] for edge deployment of LLMs in finance. [88:47] So, the strings that we're looking for [88:49] in this type of framework is [88:51] traceability, transparency, and privacy [88:53] focused. So, a lot of banks when they [88:55] work with, say, third-party companies, [88:58] they don't want to give their data away, [89:00] but what they are willing to give away [89:02] is their decision-making process. So, [89:04] what we do is we abstract that [89:05] decision-making process into a [89:07] continuous directional graph node [89:10] network. Which means that we say A leads [89:12] to B, C leads to D, and then we have a [89:15] framework that is able to connect them [89:17] together. [89:18] Okay, awesome. Hello and welcome to Make [89:20] It. So, Make It is basically helping you [89:23] get your plans out of the group chat and [89:25] actually make them happen. Um and so [89:27] this is our video demo and it plays on [89:29] the idea that basically, you know, [89:30] everybody in the group chat is super [89:32] active when they're planning something [89:33] to hang out until someone asks who is [89:35] making the reservation. And our agent [89:37] basically takes that and alleviates that [89:39] stress away from the users. And it [89:41] brings the bot right to where you are. [89:42] So, instead of having to download [89:43] another app or anything like that, you [89:45] actually just uh put it into your [89:47] iMessage. And so you can see here see in [89:49] our demo that you can add this Make It [89:51] to your iMessage group chat with your [89:53] friends or whoever it may be. [89:55] Um and it will take in preferences for [89:56] dinners or whatever event that you're [89:58] trying to plan plan with it. Um and in [90:00] the end it would actually make a [90:01] reservation for you and be able to send [90:03] it to you in a calendar invite and [90:05] actually send you reminders as well as [90:07] book your Ubers for you and things like [90:08] that, where that way all you have to [90:09] worry about is showing up. We got the [90:11] idea for this project after reading [90:13] about a fact that 40 to 70% of medical [90:16] equipment donated to the developing [90:17] world ends up being broken and unused. [90:20] That's a huge amount of equipment that [90:22] could be life-saving that is just [90:23] sitting broken and can't be fixed. And a [90:25] lot of these places there's less than [90:27] one in 100,000 people who have the [90:29] expertise to fix it. So, we decided to [90:31] create a platform using both computer [90:33] vision and access to professionals um [90:35] who know how to fix these machines in [90:37] order to help these people and teach [90:39] them how to build and fix the equipment. [90:41] So, we created this kind of like fun uh [90:43] Valentine's Day theme uh nuts and bolts. [90:46] One of the things is a lot of times [90:48] these uh manuals will be very um [90:51] vague and they can also be very outdated [90:54] and hard to understand. So, if you can [90:56] read the first step and try to identify [90:58] what the first part is. [91:00] >> the point one four six four inch with [91:04] three fourth drone dome head screw. [91:07] Okay, so [91:08] that uh the screw, dome head screw. I [91:10] have no idea what a dome head screw is, [91:11] unfortunately. This looks kind of dome [91:13] head. Is this about three fourths of an [91:16] inch? You might be on the right track. [91:18] So, when you [91:18] >> be, okay. [91:19] >> When you're ready, you'll just put it [91:20] onto um the base. [91:23] It's a bit windy out here. [91:24] >> Yeah. [91:25] Um okay. And then when you're ready, you [91:27] can click um [91:29] Step complete. And then it's going to [91:30] probably tell me like, oh, that's not [91:32] complete actually. Uh take a picture and [91:34] we'll we'll see we'll see uh how you [91:35] did. And verify. [91:39] >> And you got it? [91:40] >> Correct. Yes, but let's see the next [91:42] step. [91:43] >> okay. [91:44] All right. That one was the easy one. [91:46] Let's uh take it up a notch. [91:47] >> wing. Okay, so [91:49] is this wing the right wing? Let's see. [91:52] And is this And which way should I think [91:53] it should go in this way, right? [91:55] I have no idea. [91:57] I'm not a [91:58] person who Well, it's not really [91:59] screwing in. [92:01] Well, let's just see if it thinks I'm [92:02] right. Okay, let's try that. [92:05] Step complete. Take a photo. [92:07] And you got it. Yes. [92:09] Hi, we created Miko, which is a [92:12] real-time vision language agent [92:14] for a cell microscopy. So, the idea here [92:18] is we can actually prompt our app and [92:21] ask it to do things like cell analysis, [92:23] segmentation, [92:25] as well as like just asking like what it [92:27] actually is. So, you might say [92:30] Hey Miko, how many cells are there? [92:35] So, it can tell you how many cells are [92:36] there. [92:36] >> Yeah, approximately 25. Yeah, so what I [92:38] built is it's a fixed wing that [92:40] basically approaches the maneuverability [92:41] of a drone. [92:43] Drones really slowly become super [92:44] popular because of how maneuverable they [92:46] are, [92:47] but they suffer from a lack of speed and [92:49] lack of endurance. Yeah. So, the idea [92:50] with this is instead of having control [92:53] surface on each of the airfoils to [92:55] control it, you make the entire airfoil [92:58] maneuverable. Oh, wow. So, you can see [93:01] in each case you move the entire [93:03] airfoil, which is more efficient at [93:05] generating lift or not generating lift [93:07] when you're doing a high drag maneuver. [93:09] And then you're pushing the [93:10] maneuverability of aircraft to the like [93:12] physical limits of what you can do [93:13] rather than just what the control [93:15] surfaces can deflect. The issue with [93:16] current AI systems is that they only [93:18] work for like one device. So, like cloud [93:19] code would be on your one computer, [93:20] right? But what we want to do is work [93:22] let all your devices work together in [93:24] the same AI agent system. So, for [93:26] example, I can just talk to it like uh [93:29] Can you draw a diagram of our current [93:30] architecture? [93:32] Um and then it'll send the cloud code [93:34] and then it'll be it'll think a little [93:35] bit and then it'll send the a diagram of [93:38] our architecture to our iPad to like [93:40] make edits on. [93:42] So, yeah, the way that it works is we [93:44] have uh [93:45] certain tools that we define to create [93:47] widgets or like diagrams [93:49] um [93:50] on the iPad. So, if you've ever, you [93:52] know, walked around campus trying to [93:53] find a quiet place to study and they're [93:54] all full, [93:55] uh this is something to solve that. So, [93:57] the way we did it is we took Raspberry [93:59] Pis and used the Bluetooth sensor to [94:01] find all the Bluetooth devices around [94:03] you. And uh, if it's above a certain [94:07] number, you can set like, you know, a [94:08] capacity for a room. And if it's like [94:10] there's more than I think a couple [94:11] percent of a the room capacity, it's [94:14] like, "Oh, this is probably a quiet [94:16] place to study." We built basically mind [94:19] control robotic hinge. [94:21] Um, and that's a lot of words. But, [94:23] basically right now I'm wearing a BCI, [94:25] which is a brain computer interface. [94:27] It's reading my brain signals, my brain [94:28] waves, uh, EEGs basically. Um, it's the [94:31] electrical signals that come out of my [94:33] brain. [94:34] Um, and so right now if I relax, [94:38] my brain activity will drop. And if you [94:40] tickle me, [94:42] it goes up. Okay. Um, and so basically [94:45] we're using that data, um, these values. [94:47] See, so here it spikes to basically [94:50] control a robot, to mind control a [94:52] robot. If it detects that I'm [94:54] interested, romantically interested in [94:56] someone, uh, we're using the beta high [94:58] brainwave, uh, which is the romantic [95:00] interest brainwave. It'll basically tell [95:03] the robot to extend its arm and drop a [95:06] piece of candy in your hand. Yeah, we [95:07] trained Transformer on brainwaves uh, [95:10] associated with certain thoughts or [95:12] certain emotions. And then we use that [95:13] to then classify [95:16] then map those to robot motions. So, a [95:18] lot of the distributions that we're [95:20] familiar in life, like for example, the [95:21] Poisson distribution, the Boltzmann and [95:23] the logistic, can all be derived using [95:24] this principle for math called the [95:26] maximum entropy principle. Now, what we [95:28] did in our project is we actually asked [95:30] a large language model to come up with [95:32] assumptions and to derive these [95:33] optimization problems by itself to [95:35] eventually output the distributions that [95:37] we wanted. So, for example, we'll come [95:38] up with a natural language prompt like, [95:40] "How many cups of coffee do people drink [95:41] per day?" It'll conduct Fermi [95:43] estimation. It'll find maybe three [95:44] conflicting sources, assemble a tree, [95:46] and then like solve a Lagrangian [95:48] multiplier optimization problem [95:49] together, and then convolve the [95:51] resulting distributions so we can answer [95:53] I guess the question and come up with [95:55] like the most entropically likely most [95:57] entropically likely distribution that we [96:00] want to find. And together we built [96:02] Sprout, which is an agent deployed on a [96:05] garden bed with total actionable [96:06] observability. So what that looks like, [96:08] we take a garden bed, stuff it full of a [96:10] bunch of sensors, right? Literally we [96:12] have sensors in the soil measuring [96:13] moisture, measuring temperature, [96:15] measuring CO2 gas composition, feeding [96:17] that into an edge deployed model that we [96:19] trained on a Jetson Nano, and converting [96:21] that into actions. One thing we're [96:22] really proud of is that we actually set [96:24] up a water pump next to everything. We [96:26] know water and electricity don't mix, so [96:27] that was really hard to wire up. Um we [96:29] can actually get it to where it can [96:31] water the plants when it senses it's [96:33] dry. It can move to a plant that it [96:35] knows it's dry from the soil sensor, and [96:36] it can act on that autonomously. I think [96:38] I really like free code camp like 6 [96:39] years ago I tried to learn like React JS [96:41] or something, and like a lot of the time [96:43] like the most reliable tutorials you [96:44] guys would have like 40 minutes of like [96:46] really long of like really specific [96:48] here's how I did React staple or [96:50] something, and that really helped me get [96:51] into like web dev and like this is [96:53] different, but yeah just coding in [96:54] general. So basically what we're trying [96:56] to solve is the problem that you need to [96:58] have like 10 or more people to even do a [97:00] surgery. So our solution to this is we [97:03] can now have a primary surgeon remote [97:06] connect to our fully integrated agentic [97:09] VR platform through VR from anywhere [97:11] over the world, and this view is synced [97:13] to that camera over there, and the [97:16] doctor can now look around our entire [97:18] platform. And all few of these arms have [97:21] agents inside of them. So the doctor can [97:23] control these using natural language the [97:25] same way as he would [97:27] as he would tell assistant surgeons what [97:29] to do. We are the first model to [97:30] generate end-to-end full service to [97:32] generate music just from thoughts. So [97:35] basically we create a whole service that [97:38] takes both your is able to take your [97:39] emotions, then have while you are [97:40] listening to the music, and also do an [97:43] analysis of the representation of the [97:44] song in vector space. So we had to make [97:46] this all in house because uh they were [97:48] out of the BCI equipment. So, we 3D [97:50] printed this ourselves. And then because [97:52] the EEG supplies microvolts, we created [97:55] our own amplifier, too. [97:56] And from this EEG data, it goes into two [97:59] models. One is um an emotion classifier, [98:02] which uses a CNN using some um map of [98:05] the scalp. And then there's also an LSTI [98:08] model, which handles the music and [98:09] waves. And those go into an LLM prompt, [98:12] which creates the final um input to the [98:14] Siri API, which creates our the song [98:17] output from the EEG data. Very cool. [98:19] Think about when you're on the bus, you [98:21] know, you got a lot going on. Maybe [98:22] you're going to the airport, you have [98:23] your hands full. Um you want to do uh [98:25] something really quick, maybe send your [98:26] boss a text, maybe write a quick note in [98:28] Google Docs. Um but you know, you have [98:30] like a hundred different agents that do [98:31] all these different things. I just want [98:32] to be able to use the apps the way I do, [98:34] right? And so, let me try. Let me show [98:35] you something. [98:36] Hey Yuni, open LinkedIn and check my [98:38] notifications. [98:40] And we'll give it about 2 minutes to do [98:41] its thing. Basically, what it's doing is [98:43] it has a mobile use agent running on [98:45] this Raspberry Pi. Um it's fully it's uh [98:48] very like privacy-focused, it's open [98:50] source, actually. It's inspired by a Y [98:51] Combinator called Blue. Um what they do [98:53] usually is they hook into Oh, you can [98:55] see it's already taking action. Oh, so [98:56] it's starting to click around in your [98:57] phone. [98:57] >> Exactly. Exactly. It's going to simulate [98:58] using gestures with your fingers like [99:00] >> see if we can like showcase this to [99:01] Sure. Sure. I can just plug this in, set [99:03] this up with three clicks in my [99:04] settings. It uses any app without any [99:06] set without any login to like a million [99:08] different websites. That's the idea. Oh, [99:10] okay. So, you're setting a agent loose [99:12] on your phone to do Pretty much. [99:14] >> Yeah, and it interacts with the phone [99:15] the same way you would. Exactly. So, [99:17] this is Dive Buddy. It's a [99:19] semi-autonomous underwater submersible [99:21] for research and commercial divers. And [99:23] its purpose is to help reduce risk and [99:26] task loading by putting some of the hard [99:28] work, some of the the grunt work um onto [99:30] the robot. And so, you can focus on [99:32] finding new species and commercial [99:33] activities. And it works all through [99:35] gesture controls because when you're [99:37] underwater, you can't text, you can't [99:39] speak. So, in order to go up, you [99:41] gesture go up, and it'll go up. And in [99:43] order to go to the right, you put up [99:44] your pinky, and it should go right. Um [99:48] and you can also just go full out [99:49] straight, and then it turns on both [99:51] motors. And when you want to stop, you [99:53] stop. But you can also ask AI questions. [99:56] So, let's say you're underwater, and [99:57] you're like, "Oh, what is this species [99:58] of crab?" So, you put up two fingers cuz [100:01] that's crab in dive speak, and then [100:04] it'll ask AI [100:05] what you probably saw based off location [100:07] information. And since we're, you know, [100:09] in Stanford, um and given the marine [100:11] environments across the bay, it's it's [100:13] suggesting the invasive uh Chinese [100:14] mitten crab is probably what you saw. [100:16] Yeah, basically, this is our take on the [100:18] walker, like a very ancient piece of [100:20] technology, and it's our way to improve [100:21] people's lives. So, my grandma, she had [100:24] a surgery, so she ended up on a walker, [100:25] and she stayed on that walker the rest [100:26] of life. And it was a huge pain. Like, [100:28] she had friends who would fall, and [100:29] they'd be on the floor for 3 days. So, [100:31] what do we do? We added technology to [100:34] the walker. So, we have here you see, we [100:36] have two different pressure sensors, [100:38] detects the amount of weight you put on. [100:40] We have uh a double computer vision down [100:42] underneath. It will basically take a [100:44] video of the way you walk, and it will [100:46] count the number of steps you take. [100:47] It'll analyze your gait, whether you're [100:49] walking in the correct way, basically. [100:52] Um [100:52] overall, it's all powered through uh [100:54] ESP32. That's basically just [100:56] high-flighting all this data back to [100:59] like a an API. We run all through. We [101:02] have a whole system here where you can [101:03] see all the data that we get. This is [101:06] all off the computer model. Um we have [101:08] fall detection, which is like super [101:10] needed among old people. [101:12] And then we also have an AI agent [101:14] because we thought, well, we got to go [101:15] to our users, right? Uh [101:17] my grandmother would never be able to [101:19] operate a mobile application where it [101:20] tells her where her data is. But she's [101:21] great at talking to people. This has [101:23] been an absolutely insane experience. [101:25] There have been so many incredible demos [101:27] here that I've been able to see. Uh I [101:30] really appreciate you all coming with me [101:32] on this journey, checking out TreeHacks [101:35] 2026, seeing the hard work on display, [101:39] seeing the adoption of emerging [101:42] technologies, the application of lots of [101:45] different disciplines to build things [101:46] out. Just This is the hacker ethic. This [101:49] is the ethos right here. I've been [101:51] absolutely thrilled hanging out here [101:54] talking with all these ambitious people. [101:55] I hope you really enjoyed this. I [101:57] encourage you to go to a hackathon near [102:00] you. If you have a chance, apply for [102:02] TreeHacks 2027 and other prestigious [102:06] hackathons in your area. It is [102:08] absolutely the best way to get out of [102:11] your comfort zone, [102:13] adopt a lot of new tools, and meet a lot [102:15] of new people, and just have an insane [102:17] weekend that you'll remember for the [102:18] rest of your life. [102:20] Happy coding, everybody.