Raj Patel

The Mike Vestil Show

1. Key Themes

Human Activity Data Is the Critical Bottleneck for the Robotics Revolution

Human Archive's entire thesis rests on the idea that the next wave of robotics and world models will be bottlenecked not by compute or model architecture, but by high-fidelity, multi-modal human activity data. The company is racing to become the foundational data infrastructure layer before the market fully crystallizes.

"We're seeing there becoming a market for learning from humans. So we're just growing extremely, extremely fast." [00:16:45]

Multi-Modal Sensor Fusion Is the Technical Moat

Rather than collecting simple video, Human Archive is stacking stereo depth cameras, IMUs, wrist cameras, chest cameras, tactile gloves, and full-body motion capture into a single synchronized rig. This fidelity gap between their hardware and competitors is the core defensible advantage they are building.

"This headset captures stereo depth, it also has an IMU on the headset, and then we have two wrist cameras, we have a camera on the chest and we have two tactile gloves and then we have full upper and lower body motion capture... capturing how humans viewed the video, how humans view things, perceive depth, as well as how much force they apply on objects and how they move to complete a long horizon task." [00:11:05]

The "Get Paid More to Wear Devices" Labor Flywheel

The business model is elegantly structured: manual laborers doing tasks they already do get paid incrementally more for wearing the hardware, while the end consumer pays less than market rate for the service. This creates a self-funding, globally scalable data collection engine with natural incentive alignment across all parties.

"By wearing our devices, you'll get paid more than they're already getting paid. And for the person who's booking the service to their home, it's actually cheaper." [00:14:50]

Geographic Diversity Is Not Optional — It's a Data Quality Requirement

Raj makes a non-obvious point that collecting data only in low-cost markets would actually degrade model quality because physical environments differ meaningfully across geographies. This justifies the operational complexity of running across India, Europe, the US, and Mexico simultaneously.

"In India, a lot of homes in India don't have dishwashers, which would be a problem if your robot is in your home and doesn't know what a dishwasher is or how to interact with it. So just having as much diversity as possible is critical." [00:25:12]

Data Exclusivity Premiums as a Revenue Lever

The pricing architecture includes a meaningful exclusivity variable — frontier labs willing to pay a premium to prevent competitors from training on the same data. This creates a natural enterprise upsell mechanism and mirrors how proprietary data has been monetized in financial markets.

"Some labs will want this data where you can't share it to other labs, which means it would be exclusive, which means they would have to pay a premium for that. And then some labs would want non-exclusivity, meaning that you can send that to many other labs." [00:13:01]

Task Distribution Is Deliberately Weighted Toward Economic Deployment Probability

Human Archive is not collecting data indiscriminately. Task capture is prioritized by where robots will actually be deployed first commercially — factories and homes — rather than exotic or edge-case environments. This is a deliberate capital allocation decision disguised as data strategy.

"A lot of our task distribution is weighted heavily towards where most robots will actually be deployed economically. It's like, you know, first you're going to see robots deployed in things like factories and homes where they're going to be doing manual labor, like dishes or factory line work." [00:26:10]

Internal Model Training as Data Quality Proof

Rather than just selling raw data, Human Archive trains internally on its own datasets. This serves a dual purpose: it validates data quality to skeptical customers and opens a potential post-training data revenue stream — positioning the company upstream of model fine-tuning.

"By training on our own data, we're able to essentially prove the quality of our data and maybe even do some post-training data as well." [00:13:58]

The GPT-3 Moment for Robotics Is Imminent

Raj frames the next five years as the inflection point where a foundational robotics model equivalent to GPT-3 emerges, after which labor automation accelerates rapidly. Human Archive is explicitly positioning to be the data foundation that enables that moment.

"I think of the GPT-3 moment in robotics — within the next five years, I think it's coming up pretty soon. And I think that's going to change a lot of things, including prices will dramatically just plummet. I think manual labor will start to get automated." [00:25:12]

2. Contrarian Perspectives

Nobody — Including Frontier Labs — Actually Knows What Data Is Needed

Most investors assume frontier labs have a precise data roadmap and are simply outsourcing collection. Raj argues the opposite: even the best labs are operating with deep uncertainty about which modalities, at what scale, produce generalizable robot policies. This means Human Archive's internal research function is not overhead — it is a strategic necessity that creates proprietary knowledge the customer doesn't have.

"No one really has great hardware right now. Not even frontier labs really exactly know what data do they need to train on, what data that's scalable, what modalities to scale. No one has the answers right now." [00:31:42]

This Is an NVIDIA-Style Permanent Infrastructure Business, Not an Acqui-Hire Target

The conventional startup narrative in data businesses is: build proprietary dataset, get acquired by a large lab. Raj explicitly rejects this framing, arguing that modeling human intelligence is a permanent, multi-generational scientific and commercial infrastructure need — not a transitional stepping stone.

"We don't want to be a company that's just around for a couple of years and gets acquired. We want to set the foundation for a longer term study of humanity... I think like what we're building is not one of those things that you build for 10 years and it's not a market anymore. I think it's one of those NVIDIA-type businesses where it's going to be around for a very, very, very, very long time." [00:23:22]

The Dataset's Value in 100 Years Is Unknowable and That's the Point

Most data companies justify their assets by near-term use cases. Raj draws an analogy to the internet's creation to argue that the most important use of this data hasn't been invented yet — and that uncertainty is a feature, not a bug, of the asset they are building.

"A hundred years from now, just how when the internet was created, no one exactly knew that the internet would be used for training an LLM — just like what this dataset might even lead to, you know, the creation of the next evolution of humanity." [00:23:22]

Vertical Integration Over Partnerships Is the Only Way to Control Quality

When asked if a listener should go niche and build a dataset to be acquired by Human Archive, Raj flatly refused the premise. In a market where data quality is the product, any dependency on external collection introduces unacceptable quality risk. Full vertical control is not a preference — it is the strategy.

"We just do everything ourselves. We're going to get pure mastery over our data set by just doing that ourselves. We don't trust other people. We trust our quality of work, and we want to hold ourselves accountable for what we produce." [00:35:28]

3. Companies Identified

Human Archive

AI research lab and data infrastructure company focused on modeling human intelligence through multi-modal hardware (cameras, tactile gloves, motion capture) deployed globally to collect human activity data for robotics and world model training. Central subject of the episode; raised $8.2M seed, Y Combinator W26.

"We build different camera and sensor technology. We have an ecosystem of hardware products that we then deploy at a global scale across India, Europe, the US, Mexico, etc. And then we post-process annotate that data and conduct research on how it impacts training different robotics and world modeling applications." [00:00:00]

Figure AI

Humanoid robotics company. Mentioned as a real-world validation of Human Archive's market thesis — their robot has been running fully autonomously on an assembly line, demonstrating the commercial deployment trajectory that justifies Human Archive's task-weighting strategy.

"Figure AI, if you've seen their live stream today over the past couple of days, their robot's been working for, I think, three days, fully autonomous on assembly line." [00:26:10]

Urban Company

On-demand home services marketplace operating in India and other markets. Cited by Raj as a direct analogue to Human Archive's consumer-facing model — workers already go to homes for cleaning and other tasks, Human Archive simply layers hardware on top of that existing behavior.

"There are like many platforms that already like Urban Company is a good example of this — it's like a platform where this already exists, but they're just coming in without the hardware right now." [00:14:50]

RoboFlow

Computer vision annotation and dataset management platform. Used by Human Archive for their very first MVP — manually annotating video frames to label hands and objects in a 20-hour all-nighter session.

"We like manually annotated all of our initial data using a platform called RoboFlow. And literally from every frame in the video, like click — this is a hand or this is an object." [00:09:12]

Y Combinator

Startup accelerator. Human Archive joined W26 batch. Raj noted they had the number one most investor inbound interest on demo day, validating exceptional early market signal.

"We had the number one most investor inbound interest on Y Combinator's demo day." [00:22:27]

AWS S3

Amazon cloud storage. Used as Human Archive's primary data repository for managing thousands of SD cards worth of sensor and video data from global rigs.

"We use AWS S3. We have a center that we actually have for bulk offloading. We have servers that we've set up there that make it really quick and easy to offload." [00:34:32]

4. People Identified

Co-founder of Human Archive. Studies data science and cognitive science at UC Berkeley. Son of an Indian immigrant psychiatrist. Previously sold 16,000 mangoes and planted 100,000 trees with co-founder Shlok in high school. Led company from January 2025 founding through $8.2M seed raise.

"I studied data science and cognitive science at Berkeley. And my co-founder studied electrical engineering and mechatronics and those disciplines. So we combined that understanding." [00:06:22]

Shlok (co-founder, Human Archive)

Raj's cousin and co-founder. Top ~200 researcher in the country during high school, competed in science and engineering competitions, attended Stanford. Deeply research-focused from early age. Their fathers met in medical school in India, which is how Raj and Shlok met.

"Shlok was always focused really deeply on research. So he was a top kind of like 200 researcher in the country and he did science and engineering competitions. So he went on to go to Stanford." [00:01:43]

Samay (co-founder, Human Archive)

One of Raj's two Berkeley roommates who became co-founders. Focused on computer vision research.

"Samay is very focused on computer vision research." [00:01:43]

Ruchil (co-founder, Human Archive)

Born and raised in India. Had a previous business acquired there. Leads all global operations for Human Archive, constantly traveling between SF, India, and other markets to secure large-scale partnerships with operations-heavy businesses across Asia and the US.

"Ruchil was born and raised in India and he's acquired — his previous business was acquired there as well too. So he leads a lot of our global operations." [00:01:43]

5. Operating Insights

The All-Nighter Annotation Sprint as Customer Discovery Accelerant

Human Archive's first real customer conversations were unlocked not by a polished product but by manually annotating data in a single 20-hour session to produce a demo. The insight is that the fastest path to enterprise customer feedback in deep tech is to brute-force a credible artifact rather than wait for production infrastructure.

"We like manually annotated all of our initial data using a platform called RoboFlow. And literally that's when we go in and just from every frame in the video, click — this is a hand or this is an object. That took us like 20 hours and we just pulled a straight all-nighter for it. But our initial MVP was just some demos that we posted online." [00:09:12]

Size Your Round to Operational Reality, Not to Maximize Dilution Management

Raj describes knowing precisely how much to raise — $8.2M, not more, not less — based on the capital requirements of hardware manufacturing, global operations payouts, and research. Raising too little starves a capital-intensive model; raising too much creates pressure to deploy capital inefficiently. The discipline of target-setting before going to market resulted in hitting the number exactly.

"We knew we wanted to take around $8 million, not more than that, not less than that. And we kind of just hit our target. So things just worked out kind of perfect on that end." [00:22:27]

Keep Hardware and Software Teams Lean While Scaling Operations Headcount

Human Archive deliberately maintains a small, elite technical team while building a larger global operations workforce. This preserves technical culture and accountability while enabling the geographic scale the business requires — avoiding the common mistake of scaling headcount uniformly across all functions.

"We want to keep our hardware and software teams extremely lean as well." [00:20:33]

6. Overlooked Insights

The SD Card Offloading Problem Is a Hidden Moat

Raj briefly mentions the logistics of managing thousands of simultaneous hardware rigs each with SD cards, requiring dedicated physical servers and offloading infrastructure. This throwaway operational detail is actually significant: any competitor attempting to replicate Human Archive's scale immediately faces a physical data logistics problem that cannot be solved with software. The company has quietly built a proprietary data pipeline infrastructure — separate from their sensor hardware and ML research — that compounds their lead and that few observers would identify as a competitive advantage.

"We use SD cards for some of our hardware and, as you can imagine, when you have thousands and thousands of rigs with thousands of SD cards, offloading that and storing it is something that is difficult." [00:34:32]

EMG Sensors and Eye Tracking Are Unannounced Next-Generation Product Lines

In a single passing sentence while answering a hypothetical question about what he would do with acquired wealth, Raj reveals two specific sensor modalities — EMG (electromyography, which captures muscle electrical signals) and eye-tracking cameras — that Human Archive has not yet deployed but has clearly already evaluated. These are not idle examples; EMG data would enable Human Archive to capture neuromuscular intent (not just movement), and eye tracking would capture attentional priority during tasks. Either modality would represent a step-change in dataset fidelity and would have no comparable competitor dataset. This is a preview of their next hardware generation, hidden in a hypothetical answer.

"We haven't even gotten a chance to play around with things like EMG sensors or eye tracking cameras — there's so many different things that we can do. This is a very long horizon market." [00:33:32]