The problem
Health AI agents are starting to act, not just answer. They order labs, message patients, renew prescriptions, and update charts. Every one of those actions crosses a line where software stops reading and starts doing. Today there is no consistent human checkpoint in that loop, and audit trails are afterthoughts.
The idea
MedGate is a clinical action review console: an approval queue between an AI agent and consequential clinical actions. Nothing consequential executes until a human looks at it. The reviewer sees the agent's plan beside the untrusted instruction found inside a retrieved referral note, quoted verbatim and labeled as untrusted data, then makes one of three explicit choices: approve as requested, trim and approve, or block.
What the prototype does
- Review queue: pending clinical actions with synthetic MRNs, risk badges, and the reason each was held (off-hours, chart mismatch, untrusted text, controlled-substance ceiling).
- Case detail: the full agent plan, the untrusted note excerpt, and approve / trim & approve / block with a required decision note.
- Audit log: immutable, newest-first record of every decision: who, what, when, verdict.
- Policy settings: plain-language toggles: off-hours messaging hold, chart-mismatch holds ordering, controlled-substance ceiling, untrusted-text escalation, auto-clear routine reads.
Design decisions
- The reviewer sees evidence, not a wall of logs: why an action was held, what the agent planned, and what untrusted data influenced it.
- Untrusted content is always quoted and labeled, so the human review compares claims against provenance rather than vibes.
- Every decision is reversible in the queue but permanent in the audit log.
How we built it
A single-page front end in HTML, CSS, and vanilla JavaScript, hosted on GitHub Pages. All demo data is synthetic.
Challenges
Scope. A real deployment would need EHR integration, identity, and clinician notifications. We cut all of it to make the review surface itself excellent, and we say so plainly: MedGate is a workflow prototype, not a medical device, and it contains no real patient data.
Log in or sign up for Devpost to join the conversation.