01Trajectories
Computer-use trajectories
Human-demonstrated interactions across browser and native desktop environments, teaching models to navigate and operate across both.
How experts actually work in the digital and physical world, turned into data a model can learn from.
Backed by angels from
What we build
Human-demonstrated interactions across browser and native desktop environments, teaching models to navigate and operate across both.
Grading frameworks for reasoning and code generation, turning subjective expert judgment into scalable reward signals.
Agents learn to complete long-horizon tasks with autonomy and reliability, in isolated environments.
Physical AI
The same question, asked of the physical world: how does an expert actually do the work? We record it first‑person, where the work already happens.
First-person recordings of specialists at work — two hands, real tools, real force, long-horizon tasks. Collected inside working shops, not staged in a lab.
The vehicle, the part and the service come from the shop’s own system as the clip is recorded — not reconstructed by an annotator watching it afterwards.
LeRobot, RLDS or HDF5, with per-clip quality metrics, a data card, and a documented consent chain behind every hour.
Why we build
Today, very few labs can train an agent with RL environments and reliable data.
By 2027 that moves to the company: each organisation turning its own experts’ work into environments with verifiable rewards its agents can learn from.
We’re building for that future, and hiring:
San Francisco · Full-time
San Francisco · Full-time
San Francisco · Full-time