Evaluations and RL
Training an agent to model parts needs two things a lab can’t write without engineers: tasks that reflect real design work, and checks that can tell a correct part from a plausible one.
What we build
Automatic verifiers
Scripts that score an attempt: geometry against the reference, required features present, interference with mating parts, annotations attached and complete. We already run these on every model we deliver.
Task authoring
Design briefs at graded difficulty, from single features to multi-part assemblies, each with a reference solution and its verifier.
Benchmark sets
Held-out task sets with documented scoring, kept separate from training data.
Human grading
Engineers grade agent output where scripts can’t: design intent, manufacturability, sensible tolerances.
Failure-mode catalogues
Systematic records of where a CAD agent or copilot goes wrong, with reproducible cases and the correct result.
We work on the task, verifier and grading side. If you also need the surrounding software environment built, we work alongside your engineers or your environment partner.
Send us one task you want verified
We’ll write the verifier and a reference solution as part of a pilot, so you can see how we work on your problem before committing.