Mortgage-Lite

Mortgage-Lite is an AI-powered mortgage underwriting assistant that automates document processing, validation, and compliance checking for mortgage applications. Built with privacy-first principles, it uses a multi-agent architecture to process mortgage packets efficiently while maintaining strict data security.

Key Features

🤖 Eight Specialized AI Agents

  • Iris - Document intake and classification (all pipelines)

  • Rex - Multi-engine OCR extraction (all pipelines)

  • Val - Deterministic validation (underwriting)

  • Ana - Local LLM analysis (underwriting)

  • Claire - Cloud-based compliance checking (underwriting)

  • Servo - Servicing transfer validator (servicing)

  • Auditor - Post-close QC auditor (quality control)

  • Max - Final delivery and notifications (all pipelines)

🔒 Privacy-First Architecture

  • Local Processing: Sensitive data processed on-premises with local LLMs

  • Anonymization: PII automatically anonymized before cloud processing

  • Ephemeral Mappings: Anonymization mappings destroyed after use

  • Zero PII Exposure: Cloud services never see real personal information

⚡ Performance

  • 4x Faster: Process applications in ~4 minutes vs industry standard of 4 days

  • Cost Efficient: 60% reduction in cloud AI costs through intelligent routing

  • Parallel Processing: Multi-document concurrent extraction

  • Shift-Left Validation: Catch errors early with deterministic checks

📊 Comprehensive Features

  • Multi-Pipeline Support: Underwriting, Servicing Transfer, and Quality Control

  • Real-time Kanban Board: Track applications across pipeline stages

  • PDF Document Viewer: Inline document preview with field highlighting

  • Anomaly Detection: Severity classification (critical/warning/info)

  • Compliance Checking: Fannie Mae, FHA, VA, and Freddie Mac guidelines

  • Executive Summary: Cross-document synthesis with AI

  • Audit Trail: Complete processing history with timestamps

  • Telegram Notifications: Critical event alerts

Use Cases

1. Mortgage Underwriting

Automate the review of mortgage applications including:

  • Purchase loans

  • Refinance applications

  • FHA loans

  • VA loans

  • HELOC applications

2. Loan Servicing Transfer

Validate data integrity during loan servicing transfers:

  • Principal balance reconciliation

  • Payment history completeness

  • Escrow balance verification

  • Interest rate consistency

  • Insurance and tax certificate currency

3. Post-Close Quality Control

Audit closed loan files for defects:

  • Missing signatures detection

  • Stale appraisal checks (>120 days)

  • Income calculation verification

  • Document completeness validation

  • TRID tolerance compliance

Document Processing

Extract and validate data from:

  • W-2 forms

  • 1099 forms

  • Bank statements

  • Tax returns

  • Pay stubs

  • Appraisals

  • ID documents

Compliance Verification

Automatically check applications against:

  • Debt-to-Income (DTI) ratios

  • Loan-to-Value (LTV) ratios

  • Credit score requirements

  • Required documentation

  • Regulatory guidelines

Architecture Highlights

Underwriting Pipeline

UPLOAD → Iris → Rex → Val → Ana → [ANONYMIZE] → Claire → [DEANONYMIZE] → Max → DONE
         ↓       ↓      ↓      ↓                   ↓                         ↓
      Classify  OCR   Validate Analyze          Comply                    Deliver

Servicing Pipeline

RECEIVED → Iris → Rex → Servo → Max → TRANSFERRED
           ↓       ↓      ↓        ↓
        Classify  OCR  Reconcile Deliver

Quality Control Pipeline

SAMPLED → Iris → Rex → Auditor → Max → CLEARED/DEFECT
          ↓       ↓      ↓          ↓
       Classify  OCR   Audit     Deliver

Privacy Boundaries

  • Local Zone: Iris, Rex, Val, Ana process raw PII on local infrastructure

  • Anonymization Boundary: Data sanitized before cloud processing

  • Cloud Zone: Claire receives only anonymized data

  • Deanonymization: Results restored with real identities for delivery

Technology Stack

  • Backend: FastAPI (Python 3.12+)

  • Database: SQLite (dev) / PostgreSQL (production)

  • AI Models:

    • Claude (Anthropic) for compliance

    • Qwen 3.5 (local) for analysis

    • GLM-OCR for document extraction

  • Frontend: HTMX + TailwindCSS

  • Deployment: Docker, Kubernetes (Helm charts included)

Quick Start

See Getting Started for installation and setup instructions.

Documentation

Support

For issues, questions, or contributions, please refer to the project repository.

License

See LICENSE file in the repository root.