The challenge
Underwriters spent most of their day keying data from bank statements and ID documents, which created backlogs and inconsistent decisions.
Our solution
We built a document-processing pipeline that combines OCR, a fine-tuned extraction model and rule-based validation. Every model has an evaluation suite, and low-confidence fields go to a review queue.
Technology
- Python
- PyTorch
- pgvector
- FastAPI
- Azure