01 · What we build
Synthetic worlds, exact grades.
Training data
RL environments for physical engineering. An agent revises a design under a change memo, and every attempt earns a reward computed by physics and geometry solvers.
Evaluations
Failure reports that show exactly where a frontier model breaks: the episode, the checks it failed, the trace.
02 · Research
The future of agents is physical.
Models write production code, but they can't design a part that survives the real world. Physical AI is years behind because the training data doesn't exist. We're building it.
03 · Environments
Graded by solvers, never by judges.
Every task is graded by math, physics and geometry solvers. Load directions and the fine mesh stay hidden, and invalid geometry scores zero. Environments ship in Harbor format.
04 · Scaling and delivery
From a handful of seeds to a production pack.
Every instance is generated from a seed. A seed ships only after a reference design clears the bar and every known shortcut scores under its cap. Test seeds come from a disjoint stream, never filtered on model results.