Fintech SaaS — Invoice processing at 94% accuracy
Client was manually processing 12,000 invoices/month. We built an OCR pipeline with field-level confidence scoring and human review queue. Now 2 engineers manage what took 8.
full_breakdown() ↓
Stack: Python, Tesseract + PaddleOCR, FastAPI, Docker, PostgreSQL
Challenge: 47 distinct vendor invoice layouts, handwritten annotations, poor scan quality on 15% of documents.
Result: Field-level accuracy 94.2% on 500-doc validation set. Human review queue handles the remaining 5.8%. Processing time: 2.1 seconds per invoice average. Monthly infrastructure cost: $85 on a single t3.medium.