Doc Extractor

About this project

• •Finance teams still retype invoices into spreadsheets by hand. This does it in about 30 seconds, and — the part that matters — tells you when not to trust it.
• •Every extracted field carries a confidence score and a source citation pointing at the page it came from. In the demo it pulls vendor, invoice number, dates, totals and every line item at full confidence, then flags the currency at 67% because it had to infer USD from a dollar sign. You auto-approve the confident majority and only look at the rest.
• •Hybrid parse-then-LLM routing sends born-digital PDFs through PyMuPDF and scans through Mistral OCR, then into Claude structured outputs. A multi-signal per-field confidence engine drives model-cascade escalation when a score comes back low, and schema self-heal retries catch malformed output.
• •Extraction is validated against a 50-invoice synthetic eval set with ground truth, reporting per-field precision, recall and F1 alongside straight-through-processing accuracy. A golden-set regression test runs in CI so accuracy cannot quietly rot.
• •A FastAPI async-job backend behind a two-pane reviewer UI: PDF viewer on one side, editable field ledger on the other. Click a citation to flash its page, select PDF text to set a field, edit anything inline, and corrections PATCH straight back to the API. Deployed on Railway and Vercel.

Key Features

• •A confidence score and page citation on every field
• •Parse-then-LLM routing by document type
• •Model-cascade escalation when confidence comes back low
• •Validated against a 50-invoice ground-truth eval set
• •Golden-set regression test in CI
• •Two-pane reviewer UI built for human correction

Technologies Used