Fatoora
Invoices arrive as scans and PDFs. I built the pipeline that turns them into reliable structured data, then the analytics and checks that run on top of it.
What I built
- OCR + LLM extraction: a PaddleOCR microservice, then an LLM validation layer
- Invoice anomaly detection for automated monitoring
- LLM expense categorisation, with guardrails that stop wrong figures in generated text
- Financial analytics dashboard and AI-generated PDF reports
- React Native (Expo) mobile app with document scanning
Stack
Results
- extraction quality, internal score
- ~4 → 9 / 10
- amounts that were silently dropped, now parsed
- > 1,000 TND
- blocking issues fixed in a pre-production security audit
- 9
How it works
- Scan or PDF
- OCR · PaddleOCR
- LLM validation + guardrails
- Structured invoice
- Anomalies · dashboard · reports













