The challenge
A major insurer wanted enterprise users to interact with both structured operational data and unstructured insurance knowledge through a single governed AI experience, not another standalone RAG chatbot, but a reusable multi-agent architecture with orchestration, traceability, evaluation, governance and controlled access to enterprise data.
What we did
Contributed as senior data scientist to solution design and preliminary MVP implementation for a multi-agent platform: an unstructured knowledge agent (PDF retrieval, chunking, Vector Search, source-grounded responses, MLflow tracing), a structured data agent (natural-language analytics over governed tables, MVP on a controlled table set), a supervisor and orchestrator agent (intent routing, output aggregation, offline end-to-end evaluation), plus managed MCP tools, Unity Catalog business semantics, MLflow evaluation, agent-control patterns and a conversation UI/API specification.
The outcome
Established and demonstrated a governed enterprise agentic AI reference architecture for reasoning across documents and structured business data, forming a foundation for claims and health AI applications and validating reusable patterns for orchestration, governance, evaluation and enterprise tool integration.
Designed and preliminarily implemented a governed multi-agent architecture combining structured intelligence, supervisor orchestration and unstructured retrieval across claims and health.
Three-agent pattern with intent-based routing: Genie-backed structured Q&A, Vector Search RAG for PDFs and knowledge, supervisor aggregation with MLflow traces and evaluation, and Unity Catalog controls. Engagement covered architecture design and preliminary MVP, not full production rollout across all domains.