Prototype2025 · Co-built (small team)
Facet
A local-first pipeline that reads manufacturing drawings the way an engineer does — using vision-LLMs to locate dimensions and tolerances on the page, not just OCR them.
PythonVision LLMsPyMuPDFTesseractSQLiteApple Silicon
Overview
Engineering drawings are the hardest documents to digitize — the meaning is in where a tolerance sits, not just the text. Facet classifies a document, routes it to a specialized extractor, and uses a vision model to find the exact regions of dimensions, GD&T callouts, BOM tables and material specs, then pulls structured data with ISO/ANSI/DIN standard detection.
It runs entirely on-device, because manufacturers won't send their intellectual property to the cloud.
Highlights
- Vision-LLM coordinate detection instead of blind OCR.
- A document-type classifier routing to purpose-built pipelines.
- Ships a dashboard and an embeddable parsing widget alongside the core.