Underwriter's Assistant Chatbot
A single natural-language interface that lets insurance underwriters query cases, policies, and guidelines, replacing hours of manual cross-referencing across disconnected tools.
The Business Problem
Underwriters needed to quickly find relevant cases, compare policy terms, and check guidelines, but had to switch between multiple tools and databases for every query. Keyword search could not understand natural language or the relationships between cases, policies, and rules.
The client needed one interface where underwriters could ask complex multi-step questions and get grounded answers without manual query chaining across systems.
The Technical Solution
I built an agentic chatbot using LangGraph that orchestrates retrieval from MilvusDB (vector search) and SQL databases. The agent interprets questions, retrieves relevant context, and chains multi-step tasks (find similar cases, then fetch policy details) in one conversation.
Query refinement and task breakdown happen inside the graph, so underwriters get complete answers instead of running several manual queries across different systems.
The Scalability Factor
Deployed on AWS with Docker containers and horizontal scaling for the retrieval layer. LangGraph state management keeps conversations consistent under concurrent load.
CI/CD via GitHub Actions automates deployment of agent graph updates. Retrieval thresholds and grounding rules are version-controlled so changes are auditable and rollback-ready.
Business Impact
Underwriters query and analyze cases in natural language from a single interface.
They can query facts, spot trends, and cross-reference cases, policies, and guidelines in under 30 seconds, work that previously meant switching between multiple tools. The system scales with usage on AWS.
