What this is
An open-source generative-AI proof-of-concept we built to demonstrate that a long-form conversational AI assistant (“Clarity”) can simplify the parts of dental insurance most members find opaque — plan benefits, Explanation-of-Benefits (EoB) documents, and the underlying jargon. The member uploads a plan-benefit, brochure, or EoB PDF and chats with the assistant about what’s actually covered, what they owe, and what the form means.
It is a proof-of-concept, not a production system, but the patterns (document grounding, scope-bounded persona, safe deflection to a human) are the same ones we apply in production client work.
A live walkthrough and source review are available under NDA.
What it does
- PDF upload, instant grounding. Member uploads a plan-benefit table, brochure, or EoB as PDF; the app extracts the text and uses it as the conversation’s grounding context. A bundled example PDF is loaded if no file is uploaded.
- Persona-bounded conversation. Clarity introduces itself, simplifies dental-insurance jargon, and answers questions about what’s covered, what’s not, and what costs the member should expect.
- Stays in its lane. When asked to recommend a plan (rather than explain one), Clarity declines politely — that’s a different scope, by design.
- Tolerates messy input. The assistant is instructed to clean up typos and make conservative guesses about misspelled words, since uploaded PDFs sometimes carry OCR artifacts.
- Safe deflection to a human. When a question can’t be answered from the uploaded document, Clarity says so and offers to connect the member with a human customer-service representative — no invented coverage claims.
- Streaming response UX. Replies stream into the chat with a typing indicator, mirroring the conversational feel members expect from modern chat tools.
Built with
| Layer | Stack |
|---|---|
| Model | Leading frontier LLM |
| Frontend / UI | Streamlit (chat UI, file uploader, sidebar info panel) |
| Document parsing | PDF text extraction (open-source library) |
| Backend language | Python |
| Secrets | Environment variables |
| Response tuning | Tuned for short, focused, low-repetition replies |
What this means for you
If you operate a member-facing experience — health insurance, dental, vision, employee benefits, financial-services statements, anything where the customer routinely receives a document they don’t fully understand — we can:
- Build a document-grounded conversational AI that lets your members upload (or auto-receive) the document and get plain-language answers about it.
- Engineer persona-bounded scope so the assistant explains what’s in the document and refers everything else to the right channel.
- Add safe deflection to a human when the question is outside the assistant’s competence.
- Tune the response behavior (length, temperature, tone, repetition controls) to your member-experience standards.
Want to talk about a similar member-facing assistant in your business? Contact us.