Bringing GenAI support to millions of members

About

A major insurer needed conversational support that could answer questions about a member's own coverage, not just recite policy documents. I architected the chatbot and avatar system behind it.

Answers are produced by grounded retrieval over core business data: OpenSearch vector indexing over policy, claims and benefit records, with every response traceable back to the records it was built from. Nothing is generated without a source, because in insurance a confident wrong answer is a liability rather than a bad demo.

Alongside it runs an anti-fraud claims pipeline — Airflow orchestrating feature extraction and scoring over incoming claims to surface the ones worth a human look. Regulatory constraints kept the entire platform on-premise, so the whole thing runs on an on-prem Kubernetes cluster rather than a managed cloud service.

The client is not named here at their request.