DiPGOS
Document intelligence for construction bidding
A platform that takes a civil-construction tender from automated bid design through to progress control on site. I wrote its document intelligence layer and the retrieval behind its assistant.
- 384embedding dims
- 3chunking treatments
- 5engineers led
The constraint
A tender package is prose specifications, cost spreadsheets and CAD drawings in one folder. Chunk all three the same way and you destroy the structure that made retrieval useful in the first place.
The decision
Classify every file by title and type first, then route it to the treatment its own type needs. Embed at 384 dimensions into pgvector. Answer with hybrid keyword-and-semantic search fused by reciprocal rank — plus a calculator tool, so figures in an answer are computed rather than generated.
Where it got to
Piloted with two engineering contractors, now in pilot with Pakistan’s WAPDA for the Diamer-Bhasha Dam. Sole developer for its first two months.
Python · FastAPI · PostgreSQL · pgvector · sentence-transformers · React · deck.gl · MapLibre · Docker
