Mortgage AI¶
Multi-Agent Loan Origination -- Red Hat AI Quickstart
This reference application demonstrates how to build multi-agent AI systems on Red Hat AI / OpenShift AI using a realistic, regulated-industry business use case. This Quickstart showcases advanced AI patterns applied to the mortgage lending lifecycle, from prospect inquiry through pre-qualification, application, underwriting, and approval.
Demonstration Purposes Only
All regulatory and compliance content in this application is simulated for demonstration purposes. This is not a production-ready system and does not constitute legal or financial advice.
What is This Application?¶
The application uses a fictional mortgage lender to model the end-to-end mortgage lending process, covering:
- Prospect inquiry and pre-qualification
- Borrower application and document submission
- Loan officer pipeline management and workflow
- Underwriter review, compliance checks, and decisioning
- Executive analytics and performance monitoring
The mortgage domain was chosen because regulated financial services demand the most challenging AI patterns: role-scoped agents, compliance knowledge bases, fair lending guardrails, demographic data isolation, and comprehensive audit trails. The architecture is MVP-maturity with production structure -- component boundaries, data models, and integration patterns are designed so adopters can harden toward production without rearchitecting.
Key Personas¶
The application provides five distinct persona experiences, each with its own interface and AI agent:
| Persona | Role | Key Capabilities |
|---|---|---|
| Prospect | Anonymous visitor | Product discovery, affordability calculator, mortgage Q&A chat |
| Borrower | Authenticated applicant | Application submission, document upload, status tracking, condition responses |
| Loan Officer | Employee originator | Pipeline management, document review, borrower communication, underwriting preparation |
| Underwriter | Employee decision-maker | Application review, compliance verification, risk assessment, approval/denial decisions |
| CEO | Executive | Pipeline analytics, denial trends, loan officer performance, audit trail review |
Each persona has a dedicated chat interface powered by a role-scoped AI agent with persona-specific tools and guardrails.
AI Patterns Demonstrated¶
This Quickstart demonstrates production-ready AI patterns for regulated industries:
- Multi-Agent Orchestration: Five role-scoped agents with distinct tool sets and system prompts, coordinated via LangGraph
- Compliance Knowledge Base: RAG-powered retrieval from federal regulations (TRID, ECOA, ATR/QM, HMDA), agency guidelines (Fannie Mae, FHA), and internal policies
- Fair Lending Guardrails: ECOA compliance checks, adverse action validation, and prohibited basis detection
- HMDA Data Isolation: Demographic information stored in a dedicated schema with restricted access, isolated from general application data
- Audit Trails: Immutable append-only audit logs with cryptographic hash chains for all agent actions and decisions
- Model Routing: Rule-based routing between fast/capable models based on query complexity and tool requirements
- Safety Shields: Optional integration with Llama Guard for input/output content moderation
- Observability: Comprehensive tracing via LangFuse for agent conversations, tool calls, and model usage
Technology Stack¶
| Layer | Technology |
|---|---|
| Agent Framework | LangGraph for multi-agent orchestration |
| Observability | LangFuse (self-hosted) for tracing and monitoring |
| Model Serving | LlamaStack abstraction layer (supports OpenAI, local LLMs, OpenShift AI) |
| Backend | FastAPI with async SQLAlchemy 2.0, Pydantic 2.x validation |
| Database | PostgreSQL 16 with pgvector for embeddings |
| Frontend | React 19 with TanStack Router and Query, Tailwind CSS, shadcn/ui components |
| Identity | Keycloak (OIDC) with role-based access control |
| Storage | MinIO (S3-compatible) for document storage |
| Fairness Metrics | TrustyAI Python library for bias detection |
| Build System | Turborepo monorepo with pnpm (Node.js) and uv (Python) |
| Deployment | Helm charts for OpenShift / Kubernetes |
What's Next¶
- Architecture -- System design, component boundaries, and data flow
- Personas -- Detailed persona workflows and capabilities
- API Reference -- REST and WebSocket API documentation
Project Goals¶
This Quickstart is designed to:
- Showcase Red Hat AI / OpenShift AI capabilities through a credible, domain-rich business use case
- Demonstrate advanced AI patterns for regulated industries (compliance, audit, fairness, data isolation)
- Enable single-command local setup for rapid exploration
- Serve as a reusable template that developers can clone, deploy, and adapt to their own domain
Non-Goals¶
This is a reference application, not a production system. It does not include:
- Real external system integrations (credit bureaus, MLS, AUS, government databases)
- Real payment processing or financial transactions
- BSA/AML/KYC identity verification workflows
- Production security hardening (see maturity expectations)
- Mobile-native applications (web only)
- Multi-tenant / multi-institution support
- Automated underwriting decisions (all decisions require human confirmation)
Source Code¶
The source code is available at github.com/rh-ai-quickstart/mortgage-ai.