Case file / AI
Document Intelligence
Made every extracted table value traceable back to the exact characters on the page it came from.
- System state
- Verified
- Case ID
- DOCUMENT
Architecture map
Signal nominal
- Role
- Architecture, extraction pipeline, and viewer
- Focus
- AI systems and product infrastructure
- Stack
- Python · FastAPI · SQLAlchemy · PostgreSQL
System log
The engineering story
01 · Constraint
The challenge
OCR output is unverifiable by default: a table arrives as plain values with no way to check any of them against the document, and the cells most likely to be wrong look exactly like the ones that are right.
02 · System
The approach
Associated words to cells by area containment rather than IoU, kept both the Azure and PyMuPDF readings on every cell with the method and confidence that chose between them, and degraded to a partial record - text and coordinates intact - when structure extraction failed, so it could be re-run later.
03 · Delivery
What shipped
- 100% of populated cells resolved against the PDF text layer on the reference document
- Provenance viewer that zooms to a selected cell and flags only the exceptions worth reviewing
- Durable BullMQ-on-Postgres job queue, so extraction runs off the request path without Redis
- 260 tests across Python and TypeScript, running with no network and no Postgres