Healthcare Network Finder AI

PROBLEM

Insurance members need two things fast, in their own language: answers to coverage questions, and the nearest in-network clinic, hospital, or pharmacy that matches their plan. Support teams were doing both lookups manually — slow for the member, expensive for the operation.

MY ROLE

I designed the architecture and built the workflows: the two-agent design, the bilingual knowledge base, the provider dataset pipeline into vector search, the geolocation integration, and the WhatsApp conversational layer.

SOLUTION ARCHITECTURE

The system routes each message to one of two specialized agents rather than overloading a single prompt:

• Agent 1 — Q&A: intent classification plus RAG over an Arabic/English FAQ knowledge base for general insurance and support questions, with Pinecone vector search
• Agent 2 — Provider finder: structured search over 4,400+ provider records, filtered by specialty, network coverage, and location
• Google Places and Geocoding APIs for location-aware matching ("nearest pharmacy to me")
• Language detection so members get answers in the language they wrote in
• WhatsApp as the conversational interface, with Supabase and Google Sheets for state and data access

TECHNICAL DECISIONS

Two agents, not one. Coverage Q&A and provider search have different data, different failure modes, and different output formats. Splitting them kept each prompt small, testable, and accurate instead of one agent doing both jobs badly.

Structured data for structured questions. Provider lookup runs against the structured dataset with real filters — RAG handles the unstructured FAQ side only. Vector search is the wrong tool for "pharmacies within 5km that take my plan."

Bilingual by design, not translation. The knowledge base exists in both Arabic and English rather than machine-translating answers at runtime, which kept responses idiomatic and on-policy in both languages.

IMPACT

Provider lookup went from a manual support task to a self-service conversation, the multilingual experience improved for Arabic-first members, and the support team's lookup burden dropped. The POC validated the two-agent pattern for the client's wider AI roadmap.

WHAT I LEARNED

Location queries are deceptively hard: "near me" requires geocoding, distance logic, and graceful fallbacks when the member's location is ambiguous. The structured-data path needed as much conversational design as the RAG path.