The detail that caught me off guard is that the tool talks back. When the vision model spots an ambiguous or missing dimension on a drawing, the system generates clarification questions and the user answers them in a web form. Most conversion attempts I have seen just guess and hand you something broken, so a built-in checkpoint for uncertain dimensions seems like a genuinely thoughtful step in the pipeline.
The flow itself is straightforward to follow. You feed in a 2D engineering drawing or a hand sketch, Meta's MuseSpark vision model processes it, and the system produces named sketches and CAD features. A local web interface hosts the whole pipeline, with an interactive Three.js viewer for the resulting part and a 'Drawing vs Modeling matching table' so you can check the model against the original. The output directory holds the generated Python scripts, log files, STEP models and a FreeCAD document.
The demonstration uses a 2D mechanical drawing of an external gear with 30 teeth. The generated file opens in FreeCAD with a parametric feature tree, and when a dimension is altered the model updates and rebuilds on its own. That the sketch and feature history survive the conversion is the part worth celebrating. Files come out as .FCStd and STEP.
The maker built this because existing text-to-CAD tools produced uneditable STEP files, which is a problem I recognize. One trade-off to weigh: the project depends on an AI to run, so it is not a fully offline, self-contained converter.
Project by Vibe Coding in my Workshop. Watch it on their YouTube channel: