AI-powered CAD Document Processing Platform

The challenge

Manufacturing knowledge is spread across very different formats: specification documents, spreadsheet-based Bills of Materials, dense 2D drawings, and 3D objects. The platform needed to turn those sources into information that users could retrieve and analyze through complex queries.

What I worked on

• Built a multimodal parsing pipeline for Excel-based BOMs, PDF specifications, 2D CAD drawings, and 3D CAD objects.
• Developed a projection-based 3D object retrieval module combining computer vision, vision models, vision-language models, and RAG.
• Implemented segmentation for large 2D drawings to extract text, drawing components, legends, notes, and other structured information.
• Used traditional computer-vision methods, including flood fill, where they proved more effective than heavier model-based approaches.

System shape

The platform connected document ingestion and parsing with specialized retrieval and a multi-agent layer. Each format followed a purpose-built processing path, while the resulting information remained accessible through one query workflow.

Outcome

The result was an intelligent document-processing platform capable of working across multimodal manufacturing data rather than treating each source as an isolated file type.

Content to add: representative document examples, architecture diagram, retrieval evaluation, processing scale, and screenshots that can be shared without exposing confidential manufacturing data.

What I learned

The most advanced model is not automatically the best tool for every stage of a multimodal pipeline. Reliable systems combine learned models with simpler, well-understood algorithms when the document structure calls for them.