A production-ready healthcare support agent, built solo.

The challenge

Healthcare benefits support spans plan coverage, prior-authorization rules, and symptom guidance — each living in a different system, with a wrong answer carrying real consequences. During the AI Engineering Bootcamp, RP Wadhwa set out to build an agent that could handle all three reliably: grounded in real data, backed by real tool calls, and resilient to the failures a production system actually hits.

What he built

A multi-agent system on Google's Agent Development Kit (ADK), with dedicated agents for coverage lookup, prior-authorization checks, symptom guidance, and claims management — each backed by its own tool rather than a single do-everything prompt.

Benefits questions are answered from a live Pinecone index of real plan documentation, not a hallucinated summary — with the agent citing the source document and falling back to a clear "try again" response rather than guessing when retrieval fails. A FastAPI service layer fronts the whole system, Supabase handles session and state, and every request is traced end-to-end in Langfuse — so failures are debuggable, not just visible.

The build also handles the unglamorous parts of shipping on a live LLM API: pinned model versions, rate-limit fallbacks, and hard caps on tool-call loops so a broken tool degrades gracefully instead of hanging the agent.

The outcome

A fully working, deployable healthcare support agent — the kind of system that would normally take a team, shipped solo. It's one of the strongest examples of what an individual builder can produce during the AI Engineering Bootcamp: real retrieval, real orchestration, and real observability, not a notebook demo.

Demo