A patient-facing AI assistant that is allowed to refuse
Ask about your own health record, and watch every answer get checked before you see it.
An AI drafts answers from your record, but it never gets the last word. Safety rules run before it, and evidence checks run after it. Anything risky goes to a human clinician or is replaced with a safe, pre-written reply. Each run streams step by step so you can see who decided what.
How it works: the guardrail sandwich
The AI is the filling. Deterministic checks are the bread on both sides.
- 1
Pre-flight screenChecks that run before the AI reads your record: fixed safety rules plus an AI classifier. Emergencies and out-of-scope questions stop here.
Rules + AI classifier
Before the AI reads anything, fixed safety rules and a classifier sort your question. Emergencies and decisions only a clinician can make stop here.
- 2
Read your recordThe AI can only call read-only tools, and every call is pinned to your own patient id. A request for anyone else's record is rewritten and logged.
AI, with scoped tools
The AI looks up labs, medications and notes with read-only tools that can only ever see your own record.
- 3
Post-flight checkChecks that run after the AI drafts an answer and before you see it: every claim needs evidence, critical lab values override the draft, and an AI judge flags diagnosis or medication advice.
Code + AI judge
The draft is split into claims. Each clinical claim must rest on something actually retrieved. Critical lab values override the draft.
- 4
Publish, hold or replaceA human clinician approves, edits or declines the held draft. The decision resumes the exact paused run from its saved checkpoint.
Code decides, a clinician reviews
Code, not the AI, publishes the answer, holds it for a clinician, or swaps it for a safe pre-written reply.
Try it
Pick one of four questions, each ending a different way, or write your own.
Whose record is this?