Every Model
Describe a part in plain English and get a parametric CAD model you can edit, re-dimension, and export. The system compiles what it wrote before it shows you anything.
An LLM writing confident-looking geometry that doesn't build is the default failure mode for this problem. So the model never gets to claim success. One agent writes the CAD source, a separate agent compiles it with the real geometry kernel and diffs the result against what was asked for. If it doesn't compile, it isn't an answer.
- 01Plain English into editable CAD
- 02Compiled and verified before you see it
- 03Parametric source, not a dead mesh
- 04Seven narrow agents, not one big one
- 05Live 3D viewer in the browser
- 06Improves from its own history, with approval
In practice
You describe a part, watch it appear in a 3D viewer, then refine it by talking to it. What comes back is code with named dimensions, not a mesh. Change the wall thickness and the whole part updates. An STL is a dead end for the user; source they can own isn't.
- You describe a part→
- Router picks a path→
- Author writes the source→
- Verifier compiles and diffs→
- Tokens stream back→
- Turn lands in the journal
The agents
Seven of them, each with one narrow job. None is a general assistant and only one is allowed to emit a CAD file. Narrow agents are testable: I can assert the author emits exactly one file, and that the verifier fails closed.
- Routing a request
- Authoring the source
- Spatial reasoning
- Compile verification
- Manufacturability review
- Distilling patterns from history
- Proposing improvements for review