Skip to content
All work

Text-to-CAD Jewelry Engine

Wireframe AI

Describe a ring in plain English and get a manufacturable 3D model back — with real measurements, not just a pretty shape.

Timeline
2025 – present
Stack
React · Three.js · Gemini · Firebase · Cloud Run · OpenCascade · WebAssembly
GitHub
01Problem

  1. Problem

    Generative models make shapes you can’t cast.

    Jewelry studios turn a client’s description into a CAD model by hand. Image and mesh generators are fast, but their output has no dimensions a workshop can trust.

  2. Approach

    Predict parameters, not geometry.

    A Gemini service on Firebase Cloud Functions turns the request into about 15 design parameters — ring size, band width, prong count — with schema-enforced JSON and automatic fallback across up to six models. Values are checked against manufacturable limits before anything is built.

  3. Build

    Parametric geometry, previewed in the browser.

    A Node.js worker on Cloud Run builds watertight meshes with constructive solid geometry, queued through Cloud Tasks so heavy jobs never time out. OpenCascade compiled to WebAssembly and Three.js render the preview client-side, so 3D costs nothing on the server.

  4. Result

    Every model ships with real-world measurements.

    Stone-weight estimation error fell from 25% to under 2% across seven cuts, the LLM service is covered by 52 automated tests, and finished designs export as binary STL ready for printing.

Results