LLM4CAD: extracting information from engineering drawings
Many companies store their 2D engineering drawings only as images. Key information in title blocks, tables and dimensions is hard to search, and conventional text recognition fails on dense layouts.
We trained vision-language models on synthetically generated drawings to read this information reliably. Training ran on European high-performance computing infrastructure, while the model runs on-premises so sensitive engineering data stays in-house. Our technology partner was KIT.
- less time to extract key metadata
- −92 %
- less manual effort
- −96 %
- drawings tested
- 2,600
- GPU node hours of training
- 6,500
Source: FFplus success story