CareLoop

Inspiration

What it does

Inspiration

After a clinic visit, families remember little of what was said, and a voice assistant that guesses is dangerous: an invented instruction or a spoken dose can hurt someone. CareLoop is built so Alexa+ can only repeat what the clinician actually said.

What it does

CareLoop turns a clinician's spoken visit into a grounded care plan that families can ask Alexa+ about.

• Extractive by construction: the model returns only span offsets. Every spoken string is a verbatim clinician quote or a fixed constant, and non-verbatim quotes are rejected.
• Dose-free voice: every quantity is masked to "your prescribed amount" before speech. Full doses appear only in the authenticated companion app.
• Voice safety: crisis-to-helpline redirect, adult-only gate, prompt-injection refusal.
• Safe dialogs: destructive changes are read back and need an explicit yes. Ambiguity triggers a question, never a guess.
• Adaptive delivery: "slower / louder / softer / shorter" re-speaks the answer with no tool call or write.
• Accessibility: 48 px targets, WCAG-AA contrast, reduced motion, English/Hindi/Punjabi.

How we built it

• An MCP server with 12 tools over Streamable HTTP, served by one stdlib-only Python process, offline.
• An offline rules engine always runs. Optional Amazon Bedrock two-pass (extract, then verify).
• Four surfaces on one origin: /simulator (Alexa+ Device Lab), /card, /companion (Angular app), /judge.
• Privacy: encrypted SQLite per household, hashed transcripts, tamper-evident ledger, crypto-shred receipt.
• Auth: OAuth 2.1 with PKCE (S256), RFC 9728 metadata, rotating refresh tokens.
• The core rules are enforced in code and in CI, and re-checked independently by a preflight/ suite.

Challenges

Making "no invented advice" a property of the architecture rather than a prompt instruction. A dose leak is treated as a kill criterion, so we built a browser Device Lab and a headless driver and diffed them turn by turn to prove both devices agree.

Results (measured 2026-10-23)

• 681 Python tests and 231 frontend tests passing
• 0.0% hallucination, 0.0% dose leak, 3/3 refusals, F1 0.965
• 10 device scenarios, 221 checks, 0 dose leaks, 0 divergences
• 142 real published docs, 2,291 sentences: 0 hallucinations, 0 leaks

What's next

Real Alexa+ device testing, more languages, and clinician-side review tools. Adult care plans only; all data is synthetic.

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for CareLoop

Built With

• alexa
• alexa-plus
• amazon-bedrock
• angular.js
• astro
• docker
• github-actions
• html5
• mcp
• oauth
• python
• sqlite
• svelte
• typescript