CASE STUDY · AI & AUTOMATION · IN PROGRESS
Document Extraction with Human Review
A web app that uses AI to pull structured data out of documents and sends anything it’s unsure about to a person for review, so the output is fast and trustworthy.
FastAPI PostgreSQL pgvector React TypeScript Tailwind LLM
Problem
Teams still retype data from invoices, forms, and reports by hand. Fully automatic AI extraction is fast but makes confident mistakes, and one wrong number in a financial record costs more than the time it saved.
Approach
- Extract. An LLM reads each document and returns structured fields.
- Score. Each field gets a confidence level. High-confidence fields pass straight through.
- Review. Low-confidence fields go to a review queue where a person confirms or corrects them.
- Learn. Corrections are stored with vector search (pgvector), so similar documents improve over time.
Build status
- ✅ Two build phases complete, each shipped with passing tests
- ✅ Added a PostgreSQL test pass so database guarantees are verified on the real production database, not just SQLite
- ⏳ Next: the review queue interface and confidence routing
How I build it
I use an AI coding workflow to scaffold and test each phase, with strict rules: every phase ships with tests, and nothing counts as done until it passes. The biggest lesson so far is in the post Green Tests, Gone Wrong.