Maxicare Aria — Health Insurance AI Agent
Multi-workflow AI agent for a Maxicare + Manulife insurance brokerage — answers FB DMs, qualifies prospects, finds accredited doctors from a Postgres registry, files lead packets to agents, and runs a Stage-3 lab-result follow-up cron.
The Problem
A Maxicare + Manulife insurance brokerage was losing inquiries on Facebook — DMs about HMO coverage, accredited doctors, and lab-result follow-ups went unanswered after hours. Agents typed the same plan explainers a hundred times a day, manually searched a 5,000-row accredited-physician spreadsheet, and forgot Stage-3 client lab follow-ups. Every missed reply was a lost commission.
The Solution
Built Aria — a four-workflow AI agent system that handles the entire FB-to-policy funnel. Workflow 1 (Aria Core) verifies Meta webhooks, routes events, and runs the AI Agent grounded on a Supabase pgvector knowledge base of Maxicare + Manulife plans, with Postgres chat memory and two callable tools: find_doctor and submit_lead. Workflow 2 (KB Ingestion) lets the broker drag-and-drop policy PDFs into a form — files are chunked via RecursiveCharacterTextSplitter, embedded with OpenAI, and upserted into Supabase. Workflow 3 (Find Doctor) executes a typed Postgres query against the accredited-physician registry and formats results back to the agent. Workflow 4 (Submit Lead) sanitizes intake, inserts the lead, builds a structured email packet, and emails both the assigned agent and the patient — attaching the right application form PDF when applicable. A daily 10 AM Manila cron (Workflow 5) sweeps eligible Stage-3 leads, sends a follow-up DM via Meta Graph API, marks the row, and posts a daily summary to Ops via Gmail.
Built with Claude Code
Aria's five-workflow architecture (core agent + KB ingestion + find-doctor tool + submit-lead tool + Stage-3 cron) was scaffolded inside Claude Code with the n8n-workflow-architect agent. Tool schemas + Meta webhook verification logic validated by n8n-validation-expert before deploy; pgvector retrieval tuned with postgres-patterns.
Agents
Skills
MCP servers
Results
- Every FB DM answered in under 1 second, 24/7 — bilingual EN/TL, grounded on real Maxicare + Manulife policy docs
- Accredited-doctor lookup dropped from 3-5 minutes (manual spreadsheet) to <200 ms (typed Postgres query via tool call)
- Zero forgotten Stage-3 lab-result follow-ups since the cron shipped — daily summary lands in Ops inbox at 10:05 AM MNL
- Lead packets now arrive in the agent's inbox structured (name, plan, contact, summary, attached PDF form) — agents close instead of triaging
Gallery
Main agent workflow. Verifies Meta webhook, routes event type, extracts message, runs the AI Agent grounded on Supabase pgvector with Postgres chat memory and two tools (find_doctor, submit_lead), then sends the FB reply.