What this is
An open-source generative-AI proof-of-concept we built to demonstrate that a long-form conversational AI assistant can replace — or sit alongside — the rigid “plan compare” forms most insurance marketplaces still use. The assistant (“Pearl”) helps a dental-insurance shopper pick a plan by chatting through their location, plan-type preference, price sensitivity, and family situation, then filters a real ACA marketplace dental-plan dataset and recommends suitable plans with full coverage detail and a link back to the source brochure.
It is a proof-of-concept, not a production system — but the underlying patterns (function-calling over a structured dataset, persona-driven conversation flow, deflect-to-human safety net) are the same ones we apply in production client work.
A live walkthrough and source review are available under NDA.
What it does
- Long-form conversational intake. Pearl introduces itself, then asks the shopper for the small set of inputs needed to recommend a plan — state, county, price sensitivity, plan type (PPO / EPO / HMO / indemnity), and whether they need adult or child-only coverage. Asks for clarification when a request is ambiguous; simplifies insurance jargon as it goes.
- Function-call retrieval over a real plan dataset. Once Pearl has the inputs, it calls a structured filter function over an ACA marketplace dental-plan dataset and pulls back the top matching plans, each enriched with full coverage detail (routine, basic, and major dental care; orthodontia; deductibles; out-of-pocket maximums; premiums by household composition).
- Plan-brochure grounding. When a richer brochure exists for a plan, Pearl reads the plan brochure (markdown) and uses it to ground its answer — including a customer-service phone number and a link to the source PDF.
- Comparison and recommendation. Pearl explains the trade-offs between matching plans and recommends the best fit for the shopper’s stated situation.
- Safe deflection to a human. When a question can’t be answered from the data, Pearl says so and offers to connect the shopper with a human customer-service representative — no improvised insurance advice.
Built with
| Layer | Stack |
|---|---|
| Model | Leading frontier LLM, with function-calling |
| Frontend / UI | Streamlit |
| Data | ACA marketplace dental-plan dataset + per-plan brochure markdown |
| Data handling | Python data utilities |
| Backend language | Python |
| Secrets | Environment variables / Streamlit secrets |
What this means for you
If you operate a marketplace, broker, member-facing acquisition flow, or any product where users need to navigate a structured catalog of options through conversation rather than through a sequence of dropdown filters, we can:
- Build an AI conversational assistant over your structured catalog — health plans, dental plans, financial products, services, providers — with the same function-calling discipline so the assistant only returns options actually present in your data.
- Engineer persona-driven conversation flows that keep the assistant’s scope, tone, and escalation behavior under explicit control.
- Add document grounding (brochures, terms, EoBs) where the catalog row is not enough.
- Wire in safe handoff to a human when the question is outside the assistant’s competence.
Want to talk about a similar conversational-marketplace problem? Contact us.