Inspiration
The process of making 3D models can be time consuming and difficult. with all the new advancements in the field of artificial intelligence, we wondered if we could use that technology to make this process easier.
What it does
You type a part description, AI engineers design and test the CAD, an AI machinist finds the cheapest process, and you approve every gate.
How we built it
Models: Gemma 4 (gemma4:26b lead, gemma4:e4b sub-engineers) via OpenClaw or Ollama
- Backend: Python, FastAPI, a workflow state machine with approval gates
- CAD: CadQuery, run in an AST-screened subprocess sandbox
- Costing: deterministic model over 28 materials and 6 processes
- Frontend: [vanilla JS + three.js viewer / React + Vite]
- Integrations: Onshape API, PrusaSlicer ## Challenges we ran into
- [Small models writing broken CAD → the 3-attempt retry loop with tracebacks fed back]
- [Running model-generated code safely → sandbox]
- [Keeping parallel sub-engineers consistent]
- [Time pressure, setup problems, integration bugs]
Accomplishments that we're proud of
We managed to get from a sentence to a 3D model
What we learned
local LLM implementation, API integration
What's next for CLAD
Real Onshape push, more processes, supplier quotes, etc.
Log in or sign up for Devpost to join the conversation.