A code-compliance AI for construction engineers.
Production RAG, built to be checked.
A professional engineer needed AI answers over dense technical codes and equipment manuals that he could actually rely on: citation-backed, checkable, and covering his real document library. We built the full product: ingestion pipeline, hybrid retrieval, chat with citations, admin console, and a benchmark harness that measures answer quality instead of asserting it.
RAG demos are easy. RAG that a licensed engineer can use for compliance answers is not: the corpus is hundreds of dense PDFs (codes, standards, manufacturer equipment binders), tables matter, and a wrong answer delivered with a confident tone is worse than no answer.
So the system is built to be checked. Every answer carries citations back to the source document. A benchmark suite, with its own admin dashboard for runs and comparisons, scores retrieval quality over generated test cases, so "did the change make answers better" is a measurement, not a vibe.
What actually shipped
Upload-to-answer ingestion (storage, async status, table-aware chunking, embedding with caching); a dedicated PDF-processing sidecar for large scanned documents; hybrid vector-plus-keyword retrieval; streaming chat; template-driven document export; admin with user invites, error analytics, and the benchmark console; hardened auth with row-level security policies; rate limiting; email. Live in production with all backing services reporting healthy.
The unglamorous part that made it real
A later sprint was entirely about the client's real library, hundreds of PDFs of codes and manufacturer catalogs that broke naive parsing. Sparse-text detection, per-pattern chunker fixes, vision-model re-extraction for garbled scans. That work does not demo well and it is exactly why the thing works.