Inside the loan file: how we read 1,000 pages in minutes
A behind-the-scenes look at the document intelligence pipeline that classifies, extracts and cross-checks every page of a mortgage file before a human ever opens it.
- Written by
- LoanDNA Product TeamDocument Intelligence
- Published
- Reading time
- 8 minutes
- Topic
- Product

A typical residential loan file arrives as a single, unordered PDF: bank statements stitched to pay stubs, a 1003 wedged between appraisal photos, and a handful of pages scanned sideways. Before anyone can underwrite, audit or price that loan, someone has to make sense of it.
For most teams that someone is a person with two monitors and a checklist. For LoanDNA it is a pipeline that reads every page, decides what it is, pulls out what matters, and flags what doesn't agree.
Step one: know what you're looking at
Classification is the foundation. Our models assign each page to one of more than 400 document types, then group pages into logical documents even when they arrive out of order. A W-2 split across three scans becomes one W-2 again.
- Page-level classification with confidence scores
- Automatic splitting and re-assembly of multi-page documents
- Detection of duplicates, blanks and illegible scans
Step two: extract the data that drives decisions
Once a document is identified, field-level extraction pulls borrower names, income figures, account balances, property details and dates into a normalized schema. Every value keeps a pointer back to the exact pixel region it came from, so reviewers can verify in one click.
“The goal was never to remove the reviewer. It was to make sure the reviewer only spends time on the pages that actually need judgment.”
Step three: reconcile across the file
The most valuable signal comes from comparison. Income on the 1003 is checked against pay stubs and tax returns; the subject property address is compared across the appraisal, title and note. Discrepancies surface as exceptions with the evidence already attached.
The result is a loan file that is organized, searchable and audit-ready in minutes, with a clear list of the handful of items that still need a human decision.
