How rrbanda/mortgage-agents works
Project type and purpose
• This is a Python-based, AI-driven, full-stack backend system built on LangGraph (an LLM orchestration framework) that implements an end-to-end mortgage application workflow.
• It consists of five specialized “agents” (Application, MortgageAdvisor, Document, Appraisal, Underwriting) that coordinate to intake, guide, verify, value, and underwrite mortgage applications.
• Business rules (200+ rules) live in a Neo4j knowledge graph, and agents dispatch “tools” to perform domain-specific tasks.
File structure
• Top level:
• app/agents/… – one folder per agent, each containing an agent.py, prompts.yaml, a tools/ subfolder, and tests/
• app/utils/ – configuration loader, database connection (Neo4j), integrations (file uploads)
• graph.py & langgraph.json – LangGraph orchestration definitions
• deploy/ & mcp/deployment/ & mcp/servers/… – Kubernetes / OpenShift manifests and a standalone credit-check microservice
• Patterns:
• Modular Agent design: each agent = one node in the workflow graph
• Tools library per agent – calls into utils/database (Neo4j), external services (Twilio, file storage) or the mcp credit-check server
• CI/CD & containerization: Containerfile, build.sh, GitHub workflows, Helm/Kustomize
README for additional insights
• Workflow sequence: Customer → ApplicationAgent → MortgageAdvisorAgent → DocumentAgent → AppraisalAgent → UnderwritingAgent → Approved/Denied
• Runtime: langgraph dev on port 8123
• Production Observations: professional testing (LangSmith), compliance (Fannie/Freddie), Neo4j KG backend