When engineers need to digitize legacy drawings, they typically spend 2-8 hours per drawing manually recreating them in CAD software. We built ForgeCadNeo to automate this pipeline using AI vision and deterministic CAD geometry.
The Three-Stage Pipeline
Our system works in three stages: AI Vision Extraction using GPT-4o to read dimensions and features from scanned drawings, Parameter Validation using Pydantic schemas and geometric constraint checks to catch AI hallucinations, and Geometry Generation using OpenCASCADE's B-Rep kernel to build exact solid models.
Why B-Rep Over Mesh
The critical design decision was using B-Rep (Boundary Representation) solids instead of mesh approximations. B-Rep produces exact mathematical surfaces that import cleanly into professional CAD tools like SolidWorks, CATIA, and Fusion 360. Mesh-based approaches would require post-processing and lose dimensional accuracy.
The AI Boundary
We deliberately keep AI as a parameter extractor, never a geometry generator. The AI identifies what features exist (holes, slots, chamfers) and extracts their dimensions. The deterministic CAD kernel then builds the geometry from these validated parameters. This separation ensures reproducible, manufacturing-grade output every time.
Results
Average processing time dropped from 4+ hours of manual work to approximately 15 minutes. The STEP AP214 output passes manufacturing validation in professional CAD tools. We implemented a credit-based pricing model with Stripe for pay-per-use access.
Tech Stack
React 18 with Three.js for 3D visualization, FastAPI backend with async processing, OpenCASCADE (pythonOCC) for solid modeling, PostgreSQL with multi-tenant isolation, and Docker for deployment.