AI-Powered Insurance Recommendation Engine

CardinalSB, a US insurance agency, needed its agents to find the best plan for each client without digging through carrier documents. We built an AI-powered recommendation system on Streamlit, FastAPI, and OpenAI embeddings that went from concept to production in three months. It now serves 100+ agents across 7 integrated carriers, generating personalized quotes in seconds instead of minutes.

The problem

Carriers shared underwriting and pricing guidelines as PDFs and Word documents with inconsistent formats, terminology, and structure. Agents spent over 10 minutes per client parsing documents or checking with supervisors to find eligible plans. Without a unified eligibility engine, recommendations varied from agent to agent — lost opportunities and a degraded client experience.

What we built

A real-time plan-matching engine, embedded directly in the agency's CRM.

Fast intake

A Streamlit form captures client details — age, health conditions, location — in under 30 seconds, with validation that catches errors at entry. It needed minimal training and holds up under high agent volumes.

Document automation

Every week the system ingests updated carrier documents — PDFs, Word files, scanned guides. OCR and semantic chunking break them into structured data, which is embedded with OpenAI models and indexed in Pinecone for fast vector search across thousands of eligibility criteria.

Time-aware query generation

A custom engine converts intake responses into precise, time-aware natural-language queries. It preserves diagnosis dates, durations, and medical details, so retrieval stays accurate without brittle rule trees.

Automated delivery

Each recommendation produces a PDF summary of inputs and results, stored to a secure Google Drive folder and emailed to the agent — no screenshots or manual uploads.

CRM embedding

The tool runs inside CardinalSB's CRM via an iframe. Recommendations appear where agents already work, which meant zero training and daily usage from day one.

Stack: Streamlit + React widgets on the frontend; FastAPI on AWS ECS with autoscaling; Pinecone vector search (Weaviate compatible); OpenAI text-embedding-3-small; HIPAA-compliant architecture with full encryption and an isolated VPC. New carriers onboard in under a week.

Results

• 10× faster quoting. What took over 10 minutes now happens in under 10 seconds.
• More accurate matches. No guesswork or supervisor dependence — consistent, retrieval-backed recommendations.
• 3–4× agent productivity. Agents serve more clients, faster, with better confidence.
• Room to grow. The modular design supports CardinalSB's expansion roadmap.

"Honestly, this AI agent has been a huge win for us. Our agents get accurate quotes in seconds, without having to dig through documents or chase down supervisors. It's made their jobs easier and our client experience smoother. We couldn't be happier." — Nick Dale, CEO, CardinalSB

What's next

The roadmap includes expanding to 15+ carriers by Q4, churn modeling and premium optimization modules, a client-facing self-service portal, and computer vision that lets agents photograph medications — the model recognizes them, infers their uses, and determines plan eligibility against carrier criteria.