Chris and I are from the same hometown and have known each other for ten years, long before either of us thought about robots.

In that time I built two companies around the same problem: how do you organize people you'll never meet, in places you'll never visit, to do consistent work? The first was HireIO, a cross-border staffing platform out of Silicon Valley. The second was Talex, an expert network for LLM data annotation — around 70,000 people across ten countries. Chris went the other direction: a CS PhD at USC in computer vision, ML infrastructure at Meta, then founding ML engineer at Aven.

Running an annotation network for language models, I kept hitting the same wall from the supply side. Text was running out, and the next generation of models wasn't going to learn from the internet. It was going to learn from the physical world — and nobody had captured it. What a warehouse picker, a line cook or a phone repair technician knows is real expertise, and none of it was ever written down. You can't buy that data, because it doesn't exist yet. Chris was hitting the same wall from the model side, watching what was missing from training sets.

The default answer at the time was to put a lab in one city and record thousands of hours in the same room. That produces volume, not diversity, and robot policies don't generalize from volume. What the models needed was the opposite shape: a few hours each from thousands of different real workplaces, all over the world. Which meant the bottleneck wasn't an AI problem. It was a distributed human network problem — reaching thousands of ordinary businesses across dozens of countries, and giving the people already working in them a reason to capture what they do. That was the problem I'd spent two companies learning to solve.

We started DeepReach in 2025. Chris owns perception and the data engine; I own the network, hardware operations and the commercial side. Ten years of knowing each other means we skipped most of the negotiation that takes co-founders their first year.

Core problem: robot models can't generalize, because the data they learn from all looks the same.

Physical AI needs to see how work actually happens — a mechanic torquing a bolt, a baker shaping dough, a technician opening a phone. None of it was ever written down. The internet captured what people say about the world; it never captured how people act in it. So unlike language models, this generation can't be trained on scraped text. Every hour has to be collected on purpose.

Today the field has roughly tens of thousands of hours of this data against a need estimated in the hundreds of millions. But the gap isn't really a volume gap. It's a diversity gap. The standard approach is to build a lab, hire operators, and record for months in one controlled space. That yields thousands of hours that all look alike, and a policy trained on it fails the moment the lighting, the counter height, or the tool changes. Our customers' acceptance specs now reflect this directly: they cap how much of a delivery can come from any single environment type, and reject batches that are too concentrated. Diversity is the binding constraint, and it's now contractual.

Diversity is hard for an economic reason, not a technical one. Collecting one hour each from a thousand real workplaces means negotiating access to a thousand real businesses — each with an owner who has no reason to let a stranger film their operation, in a country where you have no presence. That cost is why everyone else builds the lab instead. It's a distribution problem wearing an AI problem's clothes.

That's why we built the company as a network rather than an operation. The people who already work in those places, and already have the trust and the access, run collection as their own local business. We give them the hardware, the software and the payments. Access stops being something we buy and becomes something the network already has.

I decided to work on this from the supply side. I'd spent two companies building cross-border human networks — a staffing platform, then a 70,000-person annotation network across ten countries — and kept watching demand arrive for data that no existing structure could produce. The bottleneck was never the model. It was that nobody had built the network that reaches the physical world.

The long-term vision is a network that can reach any workplace on earth, and a record of how human work is actually done.

Concretely: 10,000+ devices, 50 countries, one million commercial scenes, ten million task types. Every deployment adds to what we call the Human Skill Library — a permanent record of how people do physical work, built from the inside by the people who do it. Much of this knowledge exists nowhere else. It lives in the hands of a mechanic who has been doing the job for thirty years, and when he retires it's gone. Text and video on the internet captured what people say about work. Nobody has captured the work.

If we succeed, two things are different：
First, robots stop being confined to the places that can afford to build around them. Today automation only pays at Amazon scale — a facility large enough to redesign itself for the machine. That's a rounding error of the world's actual workplaces. A policy trained on a million real environments doesn't need the environment to change; it works in a repair shop with bad lighting and a counter at the wrong height. That's the difference between robotics as a capability the largest companies buy and robotics as infrastructure a small business can use.
Second, the people whose work teaches these models are inside the economy that results. The last generation of AI was trained on data scraped from people who were never asked and never paid. This one has to be collected on purpose, from real workers in real workplaces, which means it can be built the other way — as something people opt into and earn from. We already have data partners in ten countries running this as a business they own. At 10,000 operators that isn't a side effect of how we source data. It's the point.