Intelligent document processing for mortgage underwriting
Outamation, a mortgage technology company, needed to verify, classify and cross-reference borrower documents at scale. We built the core pipeline.
The problem
Thousands of applications a day, each with fifteen to thirty document types in inconsistent formats. Underwriters spent hours per file cross-referencing income, employment, assets and liabilities against lending criteria. Time-to-decision averaged five business days, and extraction errors were causing compliance issues and loan repurchase demands.
What we built
A retrieval-augmented pipeline that ingests, classifies and extracts structured data from every document, then checks the full package against configurable lending criteria. A confidence-scoring layer flags discrepancies — an income figure that differs between pay stub and tax return — for human review. SOC 2 controls and a full audit trail were built in from the start.
The result
Review time fell from five days to under forty-five minutes per application. Classification accuracy reached 97.3 percent; the system handles over two thousand applications a day with a ninety-four percent straight-through rate. Compliance-related repurchase demands fell by sixty percent.
