We'll be first to build a team to transition the current prototype into a proper V1, which we hope to sell for $750 per unit. Our initial marketing will be through events such as White Cane Day events, but also through government partnerships through organizations such as the California Department of Rehabilitation, which has programs for assistive technology lending, financing, etc. From there we would move on to wider marketing to assistive technology retail and eventually add a subscription model to incorporate maps and public transit.
Our initial team will consist of a part-time CEO and two founding engineers, one for electrical engineering, one for software, and then from there we'll add marketing and sales, customer service, and QA.
For the first year, the initial investment that we would need would be $380K. This involves the cost for 600 units of inventory and one-year salary for the CEO and engineers. The goal would be to sell all units for $450K revenue. In conclusion, the blind assistive technology market is a real market that has a real unmet need. The white cane has stayed constant for almost 100 years. But today, not only do we have powerful computer vision models, we have the advancements in edge AI hardware to run those models directly on the cane. The best time to revolutionize the white cane is today. Thank you.
So tell me a little bit about your go-to-market. Who are you going to go after first to get the early adopters for this technology?
The initial marketing strategy we were planning was first through events like White Cane Day because these are attended by many blind individuals and other assistive technology firms often do marketing there. So the plan would be to get the initial word of mouth and then from there expanding to partnering with government organizations would allow increasing the reach of the product because they have programs such as assistive technology demonstration, etc.
Does this work completely on device or does it—you mentioned that the others needed to be cloud connected. How are you doing that?
Yes. One of the main strengths of this is that it's completely encapsulated within the cane. There's no internet connection, etc., because adding internet connection requires, for example, proximity to Wi-Fi or a cell phone. So that makes it a much simpler product.
Have you talked to people who use canes and like what was the feedback they gave you?
Yes. So when I interviewed some blind individuals, one of the questions I asked was some of the difficulties they had with the white cane. I also showed them my prototype at the time and I asked for their feedback. Based on that, I developed further features such as—I noticed that many of them wanted detection of distances and directions because it helps them orient themselves in a room, etc. So this was built very much in sync with user feedback.
Any other questions?
Thank you so much.
Thank you.
All right. So, Hessian Dynamics will be presenting on their cybersecurity algorithm built to protect cyber-physical systems. Please join me in welcoming Hessian Dynamics.
That's a lot of people. Wow. All right. So, I'm Arya. I'm the co-founder of Hessian Dynamics. We're building an anomaly detection layer for autonomous systems. So, autonomous systems—what does that mean? Imagine your drone or your Waymo or your power grid or your microgrid. It's any piece of hardware that's controlled by a piece of software. It's a cyber-physical system. So as we expand to these autonomous systems becoming more and more ingrained in our lives, the Waymos, the what I just mentioned—it's going to remain imperative to maintain their operational viability and their security. And the security aspect of it, especially now, is very outdated. Traditional methods take minutes to hours to detect any sort of intrusion. And there's no real way of determining whether any fault is associated with the actual intrinsic hardware itself or if it's a malicious action. These numbers—$10.5 trillion cybercrime loss and $400 billion in OT cyber risk—are going to expand as these systems become more and more ingrained in our societies. So with that, it's going to remain imperative that we watch the physics of the system. Watch what this physics is actually doing in comparison to the command response. So that way we can distinguish between whether the hardware fault is associated with the intrinsic hardware itself or if it's actually a malicious action. Someone's taking over your system and is causing some sort of damage. And there are plenty of real-world examples of this. There was a Jeep Cherokee hack in 2015. Researchers were able to take control over some cars. Tesla hacks. There was a Ukraine power grid attack caused hundreds of millions of dollars in damage in 2015, spoofing and jamming and drones in Ukraine. It's all over the place. So, traditional IT-based monitoring and passive detection, they miss these zero-day attacks, these covert manipulations where you take advantage of unknown software vulnerabilities. So what you have to do is watch the physics of the system, compare it to what the command is telling it to do, and if there's a discrepancy in that, you can then identify what the cause of the fault was. So differentiate between fault and attack, and because you look at the physical response, as soon as something goes wrong, you can identify in milliseconds.
So this is something that we've developed over the last year or so. We validated it with the Canadian Department of National Defense. So what we do is that we have this algorithm. We create a digital twin of any system in question. It's a physics-based model. The digital twin basically is trained on the nominal data of your system, how it should operate, and then it reads the data that's in your actual autonomous system, your drone, your car, whatever. And then it compares the two—the digital twin with the actual actuator and sensor data in real time. So you detect, and then depending on the size of the residual between what the actual system is doing and the digital twin, you can actually localize whether the fault is intrinsic to the hardware or a malicious action. And then if it is a malicious action, you can correct it. You can then take the state and move it back to its original nominal state back onto the system. It's like a loop. So, and this is instant because you watch the physics.
So, our solution is control-aware. Response time is in milliseconds. We catch zero-day attacks and it works across any cyber-physical system—industrial, automotive, aerospace, robotics, energy, etc. Your traditional OT vendors, the guys who have IPOed—Darktrace, Nozomi, well Nozomi not yet, but Dragos, Claroty—they take minutes to hours only to detect and they require a human in the loop to actually correct.
So this was part of work that we validated with the Canadian DND on a $1.5 million project. We implemented it on their naval vessels, the Halifax-class naval vessel. And we were able to localize their malicious actions in milliseconds at tier five to six—eight is for drones. So your phone, for example, is at tier nine.
So with that, since incorporating last month, we raised a million dollars from South Park Commons. We are in active conversations with various pilots, including one major defense prime, and we've taken on two advisors. Reza has been with us since day zero, was the COO of Dusty Robotics, a series company in industrial robotics. Anantha was formerly the VP of autonomous driving at GM and the VP of engineering of their level five autonomous driving at Lyft. And that's my co-founder. Unfortunately, he couldn't be here today, but he got his PhD in this space. And we're both South Park Commons founder fellows. I'm a Thiel fellow and yeah, I guess I'm early but—oh, time. Okay. No time. All right, that's it. Thank you so much.
Judges, do you have any questions?
So, how does your pricing model work for events that are rare but very impactful?
So it would—so for example, let's say Waymo wanted to—so Waymo is different, their L4 autonomy, they would build this in-house. But let's say that we were talking to BMW, which is L2, L2++, and they—to build the team, and the way that these automotive vendors work is that they have two teams: they have a security team and they have a reliability team. So the security team, if they let's say required $5 million to do this, you would then charge them a discounted version of that depending on however you build it for them and then slash some X% to build it for them if they were to go to an external vendor. If you were to do this let's say on mass for a drone system, you would deploy it per drone, especially for fleets of these.
Yeah.
What is the class of problems for which this is a great way to do detection versus the ones where it's not? If that makes sense. Like where does physics-based detection really shine?
Yeah. So that's a great question. So—I took it down but we had a demo for drones. So in defense environments, there is no—I mean there are but the existing solutions are very poor. For drones, for example, spoofing and jamming are taking out fleets of drones in Ukraine. If you go online you can find that these tiny drones have miles of fiber optic cable attached to each one just to cure these spoofing and jamming attacks. So drones is perfect for it because traditional methods, if let's say you cut off—it's an RF-denied environment and you cut off the connection between ground and control, you resort to vision-based systems and those are just terrible, it doesn't work. So this is a physics-based model where you can take—you have different waypoints associated with your drone to go to a target, for example, and you would basically watch the physics of it so that way if it deviates from that waypoint, you correct it immediately back. So drone environments, to answer the question.
This field of anomaly detection is used for like parts that are not functioning and in complex systems as well. So how did you come up with this as the application versus anomaly detection in sort of production?
Yeah. So that's another good question. So when we were talking to different people in this space—robotics, automotive—they all have different needs. So watching the physics of it, differentiating between whether it's hardware fault or anomaly detection is very dependent on the industry that you're in. So when we were talking to drones, spoofing and jamming, the cyber attack aspect of it. But if you were to talk to let's say drones for commercial applications, identifying and localizing hardware faults to the exact point of hardware, that's the most important. So the broad anomaly is for both cyber attack and hardware fault depending on the industry.
I'm sorry but we're out of time. All right. Thank you so much. Thank you to Hessian Dynamics.
All right, so up next is Lipo Vision and they're working on a non-invasive real-time ultrasound imaging to better understand RNA therapeutic delivery and expression. Please welcome Lipo Vision.
Hello everyone. My name is Jiwan Yang and I'm a sixth-year soon-to-graduate student from Shapiro Lab here at Caltech. During last summer, I was working on a lipid nanoparticle formulation and delivery problem during my internship at Amgen. And it was actually through that internship that I learned about this major problem in the pharmaceutical industry, and that became the start of Lipo Vision today. So here's a one-sentence version of what we're trying to do: we make RNA delivery visible.
Let me explain why this actually matters. So about 90% of the drug candidates fail during clinical development because of three critical pain points. First, there's no real-time visibility. The teams inject the drugs and since then they're flying blind. There's no live view of where the payloads go or whether they are being expressed in the right tissue. Secondly, because of this issue, you need extra animal cohorts. To check delivery at each time point, you need to sacrifice a separate group of mice or monkeys, and that is expensive, slow, and ethically costly. Third, the toxicity shows up too late. The off-target expression is often discovered only at terminal endpoints and after you've already committed months to a formulation. And the cumulative cost of this blindness is staggering—it lets teams burn up to over $1 million on a blind iteration, not even counting the opportunity cost of delayed timelines.
So why is now the right time for Lipo Vision? There are three forces converging. First, since mRNA is a huge success, there are over 1,200 RNA drugs currently under development with more than half using LNPs as a delivery vehicle. This is a massive and growing market. Secondly, the FDA is actively pushing to reduce animal testing and teams need tools that extract more data from fewer animals, which is exactly what longitudinal imaging enables. And third, and most importantly, there is an urgent customer pull. We've done extensive customer discovery and the message was unambiguously the same: we need live expression readouts that are less complicated than what it currently is. So here's how Lipo Vision works. Today's approaches: we inject the drugs, we wait for weeks in a black box, sacrifice these animals, and do a tissue analysis. This is destructive and slow. But with Lipo Vision, we can package the gas vesicle mRNA directly inside our lipid nanoparticle. And when the LNP delivers its payload and target cells take it up, these cells express gas vesicles. These are hollow protein nanostructures that scatter ultrasound waves. Then with any standard ultrasound system, you get a live signal showing you exactly where the expression is happening—no radiation, no specialized equipment, no biopsies.
And why will customers actually care or pay for this? Well, there are three numbers that'll tell the story. We can reduce animals up to 50% fewer because we can image the same animal over time instead of having to sacrifice separate cohorts at each time point. And this will lead to reducing it to 30 to 40 days faster as expressions happen within days. And shorter validation cycles mean that we can make earlier go or no-go decisions, which is everything in pre-clinical development. And this leads to up to $200K to $1 million saved per program over 5 to 10 optimization cycles typically used in the RNA program. And our beach customers are the pre-clinical RNA delivery teams at companies like Novartis and Merck who are already working with LNPs and actively looking for better tools.
So the market in pre-clinical imaging is large, focused, and accessible, valued at up to about $3 billion. And our initial serviceable market with 10 pilot customers will give us a near-term opportunity of up to $5 to $15 million. Our revenue model starts with paid pilot studies and sponsored co-development as a CRO agent, generating revenue while we're building validation. Then we can scale up through platform licensing, implementation support, and repeat studies. This market is particularly attractive because it seamlessly fits into our customers' existing LNP workflows and ultrasound hardware. So there's nothing new to buy, just a new capability to unlock.
We've mapped the competitive landscape carefully and Lipo Vision is the only platform that is simultaneously expression-specific, real-time, non-invasive, radiation-free and safe, easy to adopt, and capable of deep tissue imaging. Our defensibility comes from a Caltech patent in gas vesicle technology licensed through OTTCP. And our long-term moat is a programmable GV biosensor platform where we can engineer these gas vesicles to light up not only at the presence but also at specific biological signals.
So where are we today? We've already confirmed the RNA-based transfection and ultrasound detection of the GV expression, and we're currently working up to LNP delivery of these mRNA with in vivo mouse studies planned next. I'm building this as a founder out of the Shapiro Lab with Professor Mikhail Shapiro as our scientific advisor, who is a pioneer in acoustic biosensing.
And here is our 12-month plan. We'll start with optimizing in vivo system leading up to one or two pilot studies with initial pharma partners, and we'll identify early customers and raise a $2 million seed round. So my ask today is for the Bill Gross prize to directly accelerate our Q1 and Q2 milestones, as this is a critical de-risking stage that turns published science into a fundable company. So in one sentence, one line: Lipo Vision makes the invisible visible so that the best RNA drugs can reach patients faster. Thank you.
Judges. Any questions?
I can start. So are you thinking localized injection or do you think it's systemic?
It's going to be—I'm imagining a systemic injection because that is the way our eventual therapeutics are getting delivered.
Yeah. But there are a lot of therapeutics that are like local injection. Why not focus on local injection because then you can actually see the distribution locally.
Initially, we can definitely use that as a kind of control to ensure that it's being expressed. But ultimately leading up to the actual RNA therapeutics, usually that gets systemically delivered and usually accumulates in liver at this point. But there are more and more targeted LNP therapeutics that attempt to deliver to different organs, and I think that's where this technology will really shine.
I don't know much about the science so I'll ask a different question. I would love to understand how you're thinking about the business model at scale. This makes sense on the pilots, it's sort of like just for the capabilities, but how does this work at success? How are you charging for it? What are you selling? What exactly is the thing you deliver and how do you charge for it?
So initially as we mentioned, it's going to be pilot studies that we're directly as a contract. But once this platform is established, we anticipate to kind of build up into a reagent that we can sell as a supply, as well as implement platform licensing, and that's going to be the way we will scale up.
Exactly.
So talk to me about the sensitivity and the relationship between the amount of RNA dose that you would receive versus the bubbles. The reason we like SPECT—not PET, but we love SPECT for pre-clinical imaging—is that you do have a linear relationship and it's very sensitive and also has excellent resolution capabilities. So contrast that a little bit to ultrasound bubbles.
Wonderful question. So there is definitely some differences in terms of sensitivity and we'll have to optimize these constructs to maximize capability, and that's going to be our Q1 phase where we're going to verify this in vitro and leading up to in vivo. It's definitely great that you mentioned the SPECT, which requires a specific facility or instrumentation and is quite expensive to run, in contrast to ultrasound where we can run things—the probe these days are less than a couple thousand dollars at most. And most hospitals and imaging facilities already own ultrasound, so we can build up on it. This is a great—but it is definitely a great point that you mentioned specificity, but it is something that we can optimize through the genetic construct optimization, which as the gas vesicles are—
Sorry, you're out of time.
Thank you, Lipo Vision.
All right, so up next we have SAI who'll be introducing a word processor that's built with built-in AI tools designed to remain under teacher guidance and control. With that, let's welcome SAI.
All right. Hello everybody and good afternoon. My name is Yash Kakaday and I'm Stephen Romero, and we're excited today to pitch you SAI.
All right. So, we're both current Caltech undergrads here. But more importantly, we both have siblings who are in high school. So the problem we're currently seeing in middle school and high schoolers is when they go to write an essay, for them to use AI, it's very easy to just finish the essay and just upload AI and submit it on their behalf. Right? Teachers can't tell what's authentic or who's written what. And this is critically ruining critical thinking skills for all students. We're not only ruining kids' critical thinking skills, but they're not developing creativity. Teachers can't tell what is written by AI or not. And this is forcing teachers for everything to be handwritten or to just block AI websites. But this isn't a real solution. Teachers are forced to make this binary choice whether to ban AI completely and make everything pencil and paper, or to let them just go free roam and just let them use AI on everything. We need a choice where students can learn how to use AI responsibly and something the teachers can trust. Enter SAI. So, we've built this word processor with built-in AI features that is controlled by teachers. So, first off, this is a fully locked-down environment in which teachers can verify that nothing's AI generated. So, students can't go ahead and paste in a bunch of AI generated content. Even if a student tries to type in a word at a time, it'll be constantly updated and checked. And if the content is seen as AI, it'll be flagged and sent to the teacher. So the teacher has some features here that they can control in addition. So students can learn how to use AI in a friendly way. So for example, you have grammar or autocomplete here. For a student, this is what it kind of looks like. Very similar to Google Docs with the AI systems on the right side.
Yeah. So we're entering a pretty big market which is the US public school system and that encompasses around 49.5 million students K through 12, but we primarily want to focus on the 26.5 million students that are in grades 6 through 12 where we see that writing volume and AI misuse risk are at its peak. We're not assuming full outreach and implementation of our tools yet into the US public school system, but our near-term goal is just around 100,000 students, which gives us a small and realistic benchmark goal to reach.
And then here's our simple business model where we see it's proven by a SaaS model as shown here with a $5 per student per year annual fee, which is similar to other educational tools such as Canvas or Gradescope.
And then here's our market adoption where we see we first want to outreach to pilots. So I want to start with a small cohort of writing-intensive teachers that create a fast, low-risk entry point into the market, and then we want to prove the integrity of our product. So basically demonstrate that students are actually submitting authentic writing work and then also log any sort of their AI usage that they're using, and then show that our environment is controlled and assisted. And then we want to scale by district. So expand from pilots, teachers to curriculum leaders for a district-wide rollout of our product. And this is proven and shown to be used in different schools to actually adopt these educational tools. All right. So let's take a quick look at what current solutions exist and what teachers and students are currently using. So in the bottom left here, we have things that don't use any assistance features and are very not secure. So we have Google Classroom, Word, and Canvas. And to be quite frankly honest, it's very easy to just go ahead and cheat and upload AI generated content in here. The only current secure solutions that exist are things like the SAT, Gradescope, and Turnitin. But here Gradescope and Turnitin don't have any necessary—they're not checking for AI constantly, it's more so just a time-based lockdown. Enter a solution in a way where we can have AI usage that we can teach students and the teachers can trust. SAI. So this brings us to competitive advantages where the teachers control—it's up to teacher discretion how much AI and what AI is allowed in a fully locked-down environment teachers can trust, where in a way that students can not only learn how to use AI but still develop those critical thinking skills that they need in today's modern world. We are the first of its kind and we've interviewed 30-plus teachers and we've built this software with teachers for other teachers. If awarded the $30K, here's exactly how we would spend it. But yeah, that's it. Thank you.
I'm going to start. So, how does it work against—why can't I just generate an essay using some AI system and then copy—like print it out and then just copy it by typing it into your system?
Yeah. So, this actually blocks you from pasting in large amounts of content. And even if you do type it in word by word, like let's say you type it in word by word, there's an AI detector that runs and it'll not only tell you that it's detected, but also flag the teacher. So this way both parties know.
So essentially at the speed of the way that it's being typed—or the actual content that's written, if it's human generated or AI generated, we can check that.
Okay. Thanks.
I'm curious as you've talked to teachers and other people if this set of functionality you have is important enough to overtake like all of the other things that a Google Docs or a Word provides for students or integrations with other systems, etc. Is it a hot enough problem for them that they're willing to switch to a totally different tool that people aren't familiar with?
Yeah, we believe so. I talked to a couple teachers and one teacher even noted 60% increase in this. And some teachers, especially English teachers, have a passion for creative writing and they really enjoy reading students' original thoughts. But when they just see a bunch of AI slop that gets uploaded to their Dropbox, it not only disappoints them, but it discourages them from making assignments and they're forced to just make them handwrite essays and it's just not a modern-day solution.
As a follow-up, how much of the other stuff—meaning like grading, turning in, like types of features—do you think you need for this to be useful? Or do you think it's just links to—like just if just the writing happens in here? I'm thinking about like everything around it, like, oh, I grade in this tool. I upload to Dropbox or to Google Classroom, like turning in assignments. There's a whole bunch of tools that are connected to this. Do you think that this can stand on its own?
Could you rephrase that question?
I guess what I'm saying is like a Google Docs and other things do a lot of other things that are integrated into tons of other software that teachers use. Like do you think you have to build all of that or do you think this can sort of stand on its own as a product?
I think we do have to build a little bit of that but it should be pretty fast. And in the current state, you can currently communicate the teacher when you can upload through this software in the current state also.
Okay maybe just a follow-up question. So like I guess challenge for you—why not just go more general and say we're going to develop a software that tells if the end product is AI generated. It could be text, it could be image, but the ability to tell that—and videos actually also too—is actually going to be very important. So, have you guys thought through that lens?
Um, yeah. Could you rephrase that question? Sorry.
So, sure. Okay. I guess—so we've all heard about fake news, right? So you see a short video on Instagram, you don't even know if it's AI generated or if it's actually a video that somebody took. Could you think of a way of actually determining content is AI generated and flag it? And the essay could be the starting point, but like if you look at the end product instead of looking at the process of how it's being written into your system, if you actually just look at the final essay, can you actually tell? Could you develop a system to tell of the final product?
Yes. Yeah. Just given the final essay, you can copy and paste and you can even try this now with AI detectors. You can check if the whole thing was AI generated or certain parts of it are human generated. Yes.
Then why not just do that instead of this system?
Because this system—we're developing a way that teachers can enable students to learn how to use AI features if it's up to the teacher's discretion. So we want to develop not only a solution that checks and prevents AI cheating, but we also don't want students to fall off the curve and we want them to learn how to use AI.
Got it. Thanks.
All right. Thank you so much you guys.
All right. Next up, we have Viger and they are developing an agentic image generation tool that's designed to integrate directly with Photoshop workflows. Please welcome Viger.
Hi everyone, my name is Suryan Aranhari. I'm currently a fourth-year PhD student here at Caltech in the Thompson Lab and I'm here to present you a project I've been working on called Viger. My research is on computer vision. For my research I've had to develop a lot of synthetic images and I'll show you some examples of this, but I found that current tools are either not good enough—and I'll show you some examples why—or are too expensive, sometimes both. Let me tell you how I propose we can solve this problem.
It should be of no surprise to you that models are only as good as the data that goes into them. This is why there are multiple billion-dollar-plus companies whose sole business plan is to make quality training data. There are other companies on the other hand that make tools to make data but cannot give this data to model trainers because of license restrictions. This is where we come in. We believe there's this immense opportunity to create tools not for people but for agents to make synthetic data.
Let me walk you through a story. A week ago, I was making some images of people playing baseball with tomatoes. Now, as you can see, I used state-of-the-art models—Narrow Banana Pro, etc., OpenAI, Image G, 4.1, 5.4, all of the above. And I found that a lot of the models fundamentally lacked understanding of baseball. They induced biases and were just not good enough. Right? And so how can we solve this problem? I argue we need quality data. Model companies, model trainers have run out of quality data by having scraped the internet and are now resorting to manual data curation. And we believe agents can solve this problem much faster, much cheaper, and in a much more controllable way.
So first we're going to make agents and all the infrastructure that goes with it for labeling and manually verifying data. We're going to create a data engine that streams this data to model trainers and acquires and puts proper licenses out. And then we're going to work with model trainers to get the specific data they need to make their models better in the directions they want to go in.
The market is massive and I don't need to convince you of this. We believe that if these models can attack as much as just video games, ads, film, TV, and graphic design, it's trillions of dollars added to the world's GDP. And we believe we can capture a significant portion of this.
The current players today are largely capitalizing on wage disparities between first and third world countries in order to use humans to manually annotate and generate data. We believe humans have better things to do and largely data production of quality data is still manual. This is why we believe we can give agents state-of-the-art tools to produce cheap and high-quality data.
Our first product is a synthetic image data generator. So while the current workflow goes from a prompt to an image, what we want to make is something that first makes a scaffold of this image, has another agent come in and find objects to fit that scaffold, another agent come in, move those objects around and find and replace objects that might be a better fit for the scaffold, another agent come in and expand this canvas to find similar objects that better fit this canvas, and a fifth agent finally come in and paint this image. As you can see, this workflow is much more regent, much more controllable, and also much more amenable to changes at each point.
Whatever I take home today will be used on key milestones to develop this demo to then raise funds at the end of the summer. We're going to build the agentic infrastructure with the key specific success metric of getting an image generator that is comparable to the state-of-the-art on this ELO benchmark of more than 1,200. We're then going to build this data agent such that the data we generate is cost-effective. And then we're going to work with model trainers to get pilots so that we can get into their training workflows.
Every dollar we take home will be used for compute, GTM, and team. With that, I'd like to thank you for your time and—sorry, this is—and while I'll take questions, I'd love for this video to play in the background of our demo.
Any questions?
I'm mesmerized by your video. Sorry. So curious. So synthetic data, what is the type of image data where this kind of synthetic generation is not a good fit versus what are the types of use cases where it's a very strong fit?
That's a great question. So one place we'd absolutely not touch is health. There's no world in which I want to make agents put a tumor in somebody's lung where it doesn't already exist. One world where we'd like to—we think it's a great fit is something like training robotics, right? If I were to—
Make an Amazon package appear on a shelf. I can very easily do that. I'm happy to give you some more examples.
Curious, there's one way to think about this business where you're in the synthetic data generation market. It's sort of you're a new kind of Mercur, some other synthetic data provider. Well, that's manual, but you know what I mean. The other is use this as a way to build the world's greatest image model. Like that's exactly why. Is that the direction you think you're headed in, or do you think this is a data play?
So that's a great question. I think there might be a great use case for trying to train our own models on this, and we believe that yes, that's actually a way in which we can convince model trainers to train on our data. And eventually, there might be a compute constraint, but I'm wondering if there's a way in which we can... because a lot of the business model for Majour and these kinds of companies is relying on people making a lot of images. We wonder if there's a more general purpose way of going about like robotics or world models. But yes.
Just a follow-up question to that, because one gap that I could see that there's a big need for is as you have more robots doing more things, you need very specific training sets to train them to do what you want them to do and not do what you don't want them to do. And right now there are humans creating all of those images. And this seems like a much more seamless way. I guess what you're hearing is, you know, if you focus in on where you can achieve the most discreet value for your company instead of working with all these other folks to enable them, you likely will have a higher value creation.
That's exactly right. Yeah, I completely agree with you. And one, the current best use case for vision models today is in robotics and world models, and we believe we can... and you know, gamedev and graphic designer, we believe we can unlock a ton of value there replacing a lot of the existing tools that do that.
So there are problems that you mentioned that like humans are generating the data for. Are there a class of problems, and I think you touched on one, where we can't generate data for? And this is like humans are not able to generate data for. So even if you hired people, that would not be helpful.
So I think this is another great question. One of the things, if you think about, right, you know, in our lab we have Ena, who is a graphic design artist, and yeah, like you know, replacing her will be like very difficult. It actually took me a colossal amount of money to make the logo for this particular presentation and the icon. And if you look carefully, you can still see off colors, right? And so there's no... I strongly believe that we want to live in a world where we can automate high value work, and you know, people, artists, graphic artists, illustrators can leverage AI tools. And if you've ever used Firefly, Adobe Firefly, it's very bad. So I think, I don't know if that answers your...
All right, thank you so much.
All right, we are powering through these and we are coming up on our second to last team. I would like to present Foundation, who will be presenting their work on an instruction set architecture design for data centers. So please join me in welcoming Foundation.
Hi everyone, thank you. My name is Vulcan Gus. I'm a PhD candidate in electrical engineering and physics here in Provis group, and also worked as a visiting researcher at Midia. I'm going to present on our essentially semiconductor IP platform that we call Foundation. The hair on fire problem that everybody probably heard millions of times about is essentially the immense scaling of AI workloads in data centers where it has led to the essential problem of power grid bottleneck. And if the current projections continue, we expect it to reach essentially the US power consumption in the next five years, which clearly will lead to tapering of this basically scaling. This led to an industry-wide basically direction to go more heterogeneous in the architectures that are being developed. So there has been a Cambrian explosion in different processor architectures as well as going more distributed sort of processors that requires more specialization to reduce the energy costs and increase the throughput of these AI workloads. You can see in the last 20 years there has been a significant increase in the last decade with the AI workloads. We basically went from one architecture to 44 architectures and this number is growing. This is where semiconductor IP is very critical. As the heterogeneity in the market increases and the dominance of a single player like Nvidia potentially decreases, the IP that's developed that will make these processor architectures compatible with data centers is going to create a lot of the value. And semiconductor IP is essentially sitting at the crossroads of two revolutions. One is AI for chip design is reducing the headcount and how these chips are being designed, as well as the increase in essentially the compute demand is increasing more players to come into market. This is basically resulting in increased demand with lower spending from a semiconductor IP business perspective. So the timing is perfect for this right now. There's significant scaling and you hear news about significant deals being made with the frontier labs and these chip makers, and a lot of the value essentially goes into the semiconductor IP that's developing these chips. Our market is basically up to a trillion market with the semiconductor sales worldwide. The semiconductor IP kind of captures like around 20 billion of that and we hope to essentially very conservatively get 1% of that, that could lead to 200 million ARR. Our customer discovery is going to be essentially targeting midsize companies that have very few chip designers in the domains that we operate in. And we have three beachhead domains where essentially improving latency and essentially reducing energy consumption is very critical, and we already have two design partners already ready to deploy and essentially tear more in conversation. Our business plan with this is with the 50 midsize chip companies we can essentially get design partnership at 5 million per deal and with a six-month tape-out cadence we can essentially get to the ARR value that we intend to get. The core expertise we have and are targeting the heterogeneous sort of computing market that is constantly growing. We believe we have the essentially best positioning to capture a lot of this emerging value and there's a significant amount of essentially core IP that we can leverage, especially in the interconnect space as well as the networking side that we can deploy. The milestones we have are clear in the sense that the first two design partnerships that we already have we want to turn into concrete deals and get to 1 million ARR in the next essentially over the summer. And we want to turn and productize these IP blocks to have essentially a portfolio and basically go for a major chipmaker partnership like Nvidia. And then eventually we can hopefully tie all of this IP together in an instruction set architecture that will essentially turn potentially this company into an ARM-like model. And we will have a domain-specific IP platform. In conclusion, we are basically leveraging the two very huge market shifts in software and semiconductors. We are differentiating our IP platform going for the heterogeneous computing approach and eventually this is sort of an emerging sort of idea but the transition from early semiconductor companies going from fab to fabless chip companies can basically make new companies to go from having individual chip designers in-house to contracting like some kind of IP platforms like ours to make designless chip companies in the future. If you have any questions, I can... thank you. Thank you for your attention.
Awesome progress. Question about design partners. So what exactly are they building? Are they super narrow specific chip designs or are they companies like Arock trying to build like a different kind of general purpose AI chip?
Yeah. So one design partner we have is essentially targeting edge AI and they're developing basically inference chips. And one specific problem that they're trying to solve is as they go into like multiple chiplet architectures, the interconnect layer of that is a big problem. So we are for example solving that problem for them. The other design partner is looking at more data center kind of developing AI inference chips for data centers. And again, the interconnect is a big problem in that space. So as we discover more and more design partners, essentially the networking layer is something that we are seeing as a significant problem that we're solving, which my PhD was on as well. So.
Just a bit about how you got there with the creation and how you bound it. So when someone's designing a chip, they have to think about where they want to send that chip, what boundary they're going to use, what are the boundaries of the current cost that they want to achieve for that chip and all these other wonderful things. So how did you build that into that design work frame? Do they basically just pull and say these are the constraints that I have, this is the cost I want to achieve it at, and then this is what I need to have as an outcome. Is it very simple and easy or do you still need a chip designer is my question.
The way we work with them is essentially they have very clear specs that they want to achieve. So they start off with they already usually have a foundry that they always work with and they have like a six-month tape-out cadence with that foundry. So we already know basically there's a process design kit from that foundry that gives us like a lot of the baseline information about what they can achieve. From then we basically meet with their team and like we understand these specs need to be achieved to meet their benchmark. And then we go in and can design, which is basically like me in this case, I can design this chip for them and then offer that solution to them. Yeah.
The existing ISA market, is it how top-heavy is it? Are you competing with like a small number of people who work in that space? Is it mostly ARM or are there a lot of companies that do this?
It is a bit top-heavy. ARM is a significant player in that space. Although ARM is targeting a lot more on the CPU side and this market shift in semiconductors is kind of changing that a little bit. Where basically this huge increase in the number of processor architectures is causing new ISAs to kind of start off. So we are hoping that we might be able to capture some of that value generation in the next five years.
So the way to think about this is this is ARM for this new kind of workload, instructor for AI workloads, GPU style stuff. Okay.
What's the biggest risk in your view for the company?
I think the biggest risk is that whatever semiconductor IP that we develop needs to be productized in the sense that you can solve customer problems but if you stay sort of incoherent across different blocks you might turn into a consulting or just like services company. So the IP blocks that we develop need to be productized to be reused across different customers. And if we can create a portfolio that gets tied together with this kind of ISA North Star, we might be able to scale things and grow significantly.
Thank you. Thank you so much.
All right, we are on our final team. Reroute Bio is developing protein circuits that selectively induce tumor cell death. Please welcome Reroute Bio.
The survival rates for many cancers have stagnated over the last 50 years. Hi, I'm Andrew and this is Lucas. Together, we're future co-founders of Reroute Medicines. And today, we're really excited to share with you our vision of creating a new class of cancer therapies that can actually eliminate cancer, not just slow them down. Oncogenes drive uncontrolled proliferation of cancer cells. Targeted therapies seek to inhibit these oncogenes with the goal of selectively inhibiting tumors. The problem is that targeted therapies are very susceptible to resistance. You can actually see this in a very simple experiment where we co-culture tumor cells in red and healthy cells in green. What you can see is that in the middle plot is that cells don't get actually killed. They just get inhibited. And this is because cancer cells are very adaptive. They can oftentimes restore other pathways and restore proliferation. So we're developing a new type of therapeutic modality called therapeutic protein circuits which seeks to directly reroute oncogenes to cell death. As you can see in this movie, circuits are highly potent, directly eliminating tumor cells in culture while sparing healthy cells, leaving no resistant colonies behind. So we deliver our proteins using mRNA LNPs. This is the same technology that was used in the COVID vaccines. So actually the drug that you see in the middle is a RAS inhibitor. RAS is by far and away the most commonly mutated human oncogene. In fact, you might have heard RAS in the news just three days ago because a breakthrough drug called Divarasib just doubled the median overall survival time of pancreatic patients from 7 months to 14 months. While this is very meaningful for patients, I think this also underscores the idea that this is only 14 months to live and resistance is inevitable. So the circuit technology that we are building is really based on decades of experience by Michael Elowitz, a professor here at Caltech and HHMI investigator who initiated the field of synthetic biology on biological circuits. We will initially develop our circuits for RAS-driven colorectal liver metastasis or short CRLM. CRLM is an ideal lead indication. First, liver delivery with established LNP chemistry is highly efficient and already derisked. Second, the median overall survival time after chemo is still less than nine months. And finally, even last line treatment alone is highly commercially viable and we actually validated the feasibility of our patient stratification strategy with diverse cohorts. Our lead circuits to tackle CRLM combine a split protein with mutant RAS-specific binders. The way this works is once delivered in healthy cells, these binders do not bind to RAS and the effectors just remain inactive. In cancer cells, as you can see on the right, that have mutant RAS, in contrast, the binders do bind to RAS. And since mutant RAS is clustered on the cell membrane, the split effectors come into close proximity. They reconstitute and actively kill the cancer cell. So over the course of the past five years and over 700 experiments, we've really built out deep expertise as well as an engineering infrastructure to design these therapeutic circuits. And so the question is, how do they even work? How well do they work? So we can test circuits in a very aggressive multifocal liver tumor model. On the left here, you can see untreated tumors from mice. Each mouse has about a dozen to hundreds of tumors that grow unchecked. On the right, what's very striking to see is that with systemically delivered LNP circuits, we can almost near completely eliminate all tumors. What about resistance? So again, we can test the circuits in a resistant model. So remember Divarasib, this is the RAS inhibitor that just doubled the overall survival for pancreatic patients. Well, in this very challenging, widely validated model of resistance, Divarasib has basically no efficacy at all. While circuits, it retains excellent potency. Finally, how generalizable are these results? What we can see is across diverse tumor types and also mechanisms of resistance to inhibitors, we can see that for drugs, they don't achieve meaningful elimination of these cells, while circuits can achieve near complete elimination of the same cells. Our goal is to raise seed financing at the end of the year. And there are two key things that are still missing for that. First, we want to validate our circuits explicitly in a mouse model of CRLM, our lead indication. And then secondly, Bill Gross funding could uniquely enable essential non-academic work related to incorporation, immigration, IP, etc. The seed funds itself will be used to advance our CRLM circuits to IND stage and we have also detailed milestones for that. So let's summarize: circuits are as potent as chemotherapies, as specific as targeted therapies. In contrast to cell and immunotherapies, they can both access intracellular targets and are also cheap to manufacture. But most importantly, circuits have the potential to finally overcome resistance. Looking ahead, we've designed a development path that enables us to stepwise derisk our platform. And our ultimate dream is to develop a single off-the-shelf pan-cancer circuit that can target and eliminate the vast majority of all human cancers.
So what we always come back to with these solutions with circuits is how can you get a sufficient amount of delivery to enough cells to make a difference. And I think your in vivo data is very promising but talk a little bit more about how the field has evolved and how about local administration is going to help you get there.
Okay. So yeah, that's exactly the still the biggest risk with this technology. And I think there are really two things to mention here. So first of all, I mean the field of LNP delivery has massively improved and there are a lot of people working on this. And the reason we chose liver as a first indication is because you can already get like basically like very high penetrance in these model systems. The second thing is actually our circuits can trigger very different outputs and one of the outputs that they can trigger is pyroptotic cell death which is a mode of cell death that can lead to active immune recruitment and literature has been shown that you can basically with this eliminate cancer cells that didn't even receive your circuits. And then one more thing is that actually we know this is a giant problem for us and so we are very actively looking out and talking to companies about this. Just earlier today actually at noon we talked to the CEO of Genevant as you know is the leading LNP provider and so we are very aware of this issue.
Can you talk me through again the design? So you have a split protein detecting mutant RAS. I'm missing the circuit part of this. So okay, how is the circuit going to help you as this cancer cell mutates and develops additional mutations? Is the single circuit supposed to be able to capture those evading cells or like walk me through the logic of this?
Yeah, so this is actually a really fundamental difference between drugs and circuits. So with drugs, all it takes is a single copy of a drug target that can mutate away from the drug for you to get resistance. Because we're not inhibiting RAS, we're sensing RAS. You actually have to mutate away every single copy of RAS for us to lose our sensitivity. And what's actually really important to know is in these RAS-driven cancers, they oftentimes amplify their gene copy. And so you'll have about 30 to 50 copies of RAS, the odds of every single copy being mutated away is very low.
So the active sensor is picking up different mutant RAS and so...
Yes, we designed them to be broadly specific to all like more majorly clinically relevant oncogenic RAS mutations. Yeah.
Okay. But then it's a split protein. Where is the circuit part of it? I don't see a circuit in here.
Yeah. So that's actually an interesting thing. Initially these circuits were developed with multiple proteins and just to make them more clinically relevant we have condensed this actually down. And we can encode this in a single protein. And the effector part here is like a split caspase that basically gets reconstituted and then executes cell death.
Right. So you're essentially delivering a protein that's going to be expressed. Okay. Thanks.
Question. Okay. Thank you so much. Thank you.
With that, I would like to thank all of our finalists for their thoughtful, impressive presentations. You all did a wonderful job. So, thank you very much. Next, you guys will... the audience will have the opportunity to vote for your favorite team and you'll watch the first video in a series highlighting innovations that are coming out of Caltech and JPL. With that, I would like to invite the judges to follow me to the deliberation room. The winners will be announced in a little bit in the breezeway during the reception. And at this time, I would also like to invite Brianna Chu from our office who will be leading this portion of the event. So, thank you all.
Hello. Hello everyone. Can we just give all of our finalists one last round of applause? Come on now. Thank you all so much. Thank you everybody who was able to make it here right here in Beckman 100 and everybody who is watching online. I want to invite all of you to take out your phones if you would like to vote for your fan favorite. If you were here last year, you'll know that we introduced this fan favorite award so that you guys who were watching could also weigh in on your particular favorite team. So, let me just give you all a moment to... So, while you vote, we have a very special sneak preview of a video series that our office has been working on. If you're a longtime friend of the office, you already know what tech transfer is. Likely you know how our office is the bridge between technology that happens in the lab and technologies that we use every day and maybe can't imagine living without. But we know that not everyone is as familiar with tech transfer as we are. I certainly didn't know anything about tech transfer when I actually joined the office of tech transfer. Maybe a little bit of a problem. But we wanted to work on bringing the stories of the Caltech and JPL inventors and researchers to you so that you get the story behind how some of these amazing, incredible inventions that you may not even know are Caltech and JPL technologies came to be and what happened to get them there. And one of those is the phone that you just used to scan this QR code. So without further ado, we would love to show you this video about the CMOS image sensor.
