Document-AI underwriting pipeline for a regional US bank

The bank ran underwriting on manual reading. Borrowers sent tax forms, balance sheets, rent rolls, and statements as scanned PDFs in any layout, and underwriters hunted for fields, retyped them, and chased what was missing. A typical file took 15–30 minutes. The hard ones took 6–8 hours, at a volume of 1,000+ applications a month.

The architectural bet was a universal parser instead of a template library. Per-template rules break the day a borrower sends a new layout, so we built extraction that reads all 11 document types in any format, scanned, rotated, or structured, with no hardcoded field positions. Confidence is scored per field, and anything the parser is unsure of goes to a reviewer rather than into the template.

Risk stayed contained because the pipeline never makes the credit call. Extracted data auto-fills the bank's own Excel underwriting template, color-coded by source and completeness, and the system flags missing fields and inconsistencies before a decision is made. Gaps surface at intake, while there is still time to go back to the borrower.

What the bank owns now is a document intake layer for its lending operation. Staff drag a whole bundle into a web app, the system processes it in parallel, and a ready-to-review package comes back in seconds. That returns ~400+ staff-hours every month and puts underwriter attention on the credit decision itself.