AI Bank Statement Processing Pipeline

Problem

NBFCs needed to analyse bank statements from 40+ banks — a process that took 2–3 days per applicant.

Challenges

• 40+ bank formats with inconsistent layouts and scanned PDFs
• Need EMI, suspense, and cash-flow signals for credit decisions
• Latency target under 60 seconds end-to-end for underwriting UX
• Strict validation before any report reaches credit officers

Solution

Multi-bank OCR pipeline that parses any format, categorises transactions, detects suspense entries, calculates EMI obligations, and generates a credit analysis report in under 60 seconds.

Architecture

PaddleOCR normalises pages; LLM extraction maps transactions to a schema; Pydantic validates amounts/dates; Redis caches repeated statements; FastAPI serves credit reports on Azure with Dockerized workers.

Timeline

4–8 weeks typical for similar builds

Lessons learned

• Schema validation after LLM extraction is non-negotiable for lending
• Bank-format edge cases need a living fixture set, not one-off prompts
• Cache identical uploads to protect cost and latency