Case study · AI engineering · Text-to-CAD
NL-CAD
Plain English to CAD models: 46/46 eval cases pass at about $0.01 each.
Describe a mechanical part in plain English and get a parametric CAD model back: STEP and STL files, the CadQuery script, and the plan it compiled from.
- Type
- Open-source AI engineering project
- Built by
- Maarij Aqeel
The challenge
LLMs that write CAD code directly fail in two ways: scripts that crash, and scripts that run fine but build the wrong part.
What I built
- The LLM writes a validated, typed plan (an OpGraph IR) instead of code. A deterministic compiler turns it into CadQuery, so syntax errors are structurally impossible.
- Geometric validators check bounding boxes and volumes to catch parts that run cleanly but are wrong.
- Vague requests get clarifying questions instead of guesses.
- A pool of warm CadQuery workers, so dimension sliders rebuild instantly with no model call.
- Edits like "make the holes 8 mm" go through a cheaper model as a one-key patch.
Results
- 46/46eval cases pass, at about $0.011 each
- ~48 msper build on a warm worker (vs ~2.6 s cold)
- Real CAD output: STEP files open in FreeCAD
Tools
- Python
- CadQuery
- FastAPI
- OpenRouter
Want something similar? See ai product builds →
More case studies
Voice AI · MelonLabs
WordBuddy AI
Shipped to Google Play with 80% lower conversational-AI costs.
Read the case study
AI automation · 16x9.ai
AI Content Automation Engine
Cut content production time 95% with four chained n8n workflows.
Read the case study
Voice AI · Web app
Intervue AI
Real-time AI interviews with 80% lower voice costs and 30% faster setup.
Read the case study