Featured case study
Accounts payable automation where AI extracts, and code decides
At-a-glance invoice summary
AI extraction, with confidence scores
Deterministic validation rules
Full audit trail
An AI accounts payable pipeline where the model's only job is reading the invoice. Every decision that touches money runs through separate, deterministic, unit-tested code the model never touches.
The problem
Most "AI invoice automation" hands an LLM the whole decision, including what gets approved for payment. That's the part that should worry a finance team: a model that can misread a field can just as easily misjudge a policy. I wanted to build the safer architecture and prove it against real invoices, not a slide deck.
What I built
- A full pipeline: text extraction → AI field extraction with confidence scores → deterministic vendor & PO matching → 7 hand-written validation rules → deterministic approval decision → human review → signed accounting export
- A state machine and append-only audit log, so every status change is attributable and explainable after the fact
- An eval harness measuring extraction accuracy and discrepancy-detection precision/recall against real supplier invoices, not just a working demo
- CI, migrations, and a deployed instance, treated like production software throughout
Stack
Python
FastAPI
PostgreSQL
SQLAlchemy
Celery
Docker
Railway
pytest
100%
field extraction accuracy on real invoices
100%
precision & recall on discrepancy detection
0
financial decisions made by the AI model
7
deterministic validation rules, fully tested
