The problem
Assessing one client meant reading a stack of scans by hand: the loan agreement, payment schedules, annexes, a BIK report — mixed formats, mixed scan quality. Pulling out the numbers, reconstructing the schedule and estimating what the client could actually gain took hours per case and gated how fast an offer could go out.
At ~50 agreements a month, the manual read was the bottleneck of the whole practice. And plain AI wasn't an answer the firm could accept — a lawyer can't put an unverifiable number in front of a client.
What I built
A hybrid OCR (EasyOCR/Tesseract) + LLM pipeline that extracts every parameter of the loan — with a full audit trail. Each field carries its value, the source document, the page number, the exact text snippet and a confidence score, so a human verifies any number in seconds instead of re-reading the file.
On top of extraction: payment-schedule reconstruction, incurred and future cost analysis, early-repayment and refinancing benefits, commission and insurance refunds, and flagging of legal catches and potentially abusive clauses. The frontend serves a summary of the agreement — legal risks and the financial picture — ready to turn into a client offer the same day.
The documents are sensitive financial data, so the whole system — frontend, Python backend and the AI-model API — runs on-premise at the firm. Nothing leaves their infrastructure (GDPR).
Stack
Python · EasyOCR · Tesseract · LLM · OCR pipeline · On-premise