Problem
Replacing a broken or missing part often starts with nothing more than a few photos, a sketch or some caliper readings. Turning those into an accurate, editable CAD model normally requires fluency in a CAD program. Vision LLMs describe images well but read measuring instruments unreliably.
Approach
RaisingImages lets people who are not fluent in FreeCAD create parametric 3D models. A vision LLM works out what each image shows. Drawings and grid-paper sketches can give a rough first draft. Caliper photos are read by an OpenCV reader and by repeated vision reads, and each value keeps its evidence. Each photo is shown next to its measurement fields, so the operator can check and correct a value against the image it came from. An LLM writes a parametric FreeCAD script, which runs headless and is repaired automatically when it fails. The operator answers questions in a chat and can confirm or override any value.
Results and Learnings
It is not always quick: unclear photos, ambiguous sketches or a failing build can take several rounds of questions and corrections.
Seeing each photo side by side with its measurement input helped most with reverse engineering.
LLMs are good at finding things in images and writing CAD code. Measurements still need deterministic checks and a person who decides.
Next Steps
Make the pipeline less fragile.
Make the UI smoother.
Further debugging.
Improve vernier caliper reading.
Increase speed / reduce wait
