An AI Underwriting Pipeline That Turns Documents Into Decision-Ready Insights
Devine Globe combined generative AI development, intelligent document processing, financial validation, and serverless cloud engineering to create an automated mortgage underwriting workflow for income, expense, loan, and eligibility analysis.
1. Intelligent Mortgage Document Processing
We created a document-processing pipeline that accepts payslips, salary certificates, bank statements, loan agreements, tax forms, and supporting financial records. The system classifies each document and routes it to the appropriate analysis module, reducing the need for manual sorting and data entry.
2. Automated Income Analysis and Verification
AWS Textract performs optical character recognition, while Amazon Bedrock with Nova Pro extracts and interprets income information. The module compares payslips with salary certificates, detects bonus payments, checks salary consistency, and processes self-employed income and Form 11 tax documents.
3. AI-Powered Expense and Loan Analysis
GCP Gemini 2.5 Pro analyses bank statements to identify recurring expenses, loan repayments, unusual transactions, and existing financial commitments. EMI obligations are extracted from loan documents and cross-validated against bank activity and declared income.
4. Eligibility Scoring and Decision Support
The Summary Analysis module combines verified income, expense, and loan data to calculate the applicant’s debt-to-income ratio. It generates an eligibility score from 0 to 100 and presents an Approve, Conditional, or Reject recommendation with detailed reasoning, confidence indicators, and suggested actions for the underwriter.
5. Scalable Serverless Mortgage Processing
The AI pipeline runs through seven sequential AWS Lambda functions, allowing applications to be processed without relying on fixed infrastructure. An AWS Amplify interface enables underwriters to upload files, follow processing progress, review queries, examine flagged inconsistencies, and access final eligibility results from one dashboard.
.2qinst50l_1uj.png)
