CareCompass

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for CareCompass

Inspiration

Medical bills are confusing and people ask about them out loud: "What do I actually owe?" "Can I pay over time?" "I keep missing appointments, who can help?" Often the person asking is an adult child helping a parent. LLMs are fluent but get numbers wrong, and a wrong number on a medical bill does real harm. (No statistics are claimed in this writeup.)

What it does

• estimate_patient_bill: deterministic deductible, coinsurance, copay and out-of-pocket math, plus a short spoken answer
• payment_plan: 1–24 equal interest-free payments
• screen_care_barriers: transport, cost, language, caregiving, housing, literacy and phone-access screener that returns actions plus verified resources (decision support only)
• caregiver_recap (new in v0.2): keeps state across turns in the same MCP session (Mcp-Session-Id) and reads back a recap of the bill, plan and next steps for a family caregiver. It repeats only stored tool results, and sessions don't leak into each other (tested).
• An Alexa+ simulation: a free LLM orchestrates the tools over multiple turns in a single utterance (bill → payment plan), and a grounding guard blocks any dollar amount the tools didn't produce.

How we built it

Python stdlib MCP server (JSON-RPC: initialize, tools/list, tools/call, ping; Mcp-Session-Id; readOnlyHint annotations; structuredContent). Built on two open projects by the same author: BillPilot (billing math with 24 evals) and SightBridge (screener with 20 evals). Agent loop: OpenRouter free models (Nemotron 3 Super; Gemma 4 as fallback) using OpenAI-style function calling.

Challenges

The first real run found a bug: with an untyped answers object the LLM invented keys (transport, caregiving) and the screener silently returned "low risk." Fix: a strict JSON Schema with additionalProperties: false, plus the server now reports unknown_answer_keys.

Accomplishments

Every number the user hears is computed or verified by code. 7/7 MCP HTTP tests, 2/2 session-recap tests and 3/3 grounding tests pass, and real end-to-end runs are saved in demo/.

What's next

Deploy on a free host for a real Alexa+ connection, add an MCP Apps card with an itemized bill, and add Spanish voice support.

Built With

• gemma
• json-rpc
• mcp
• nemotron
• openrouter
• python