Inspiration
Seniors in assisted living often manage multiple chronic conditions, and Chronic Care Management (CCM) depends on frequent check-ins to catch problems early. But care facilities are stretched thin, with staffing shortages around [23%], so daily check-ins often don't happen.
AI companions could fill that gap, but there's a catch: to be useful, an agent needs access to sensitive health records. Facilities can't hand that access to an AI they can't see, limit, or shut off. We wanted to build both halves: AI that checks on seniors, and access controls that check on the AI. That's why we called it HumanLoop. A human always stays in control.
What it does
HumanLoop has two parts:
- User App (Senior): a voice-first AI companion that runs daily health check-ins. It asks guideline-based symptom questions, accepts photos (like a swollen ankle), reads connected device data (like a scale), and alerts the nurse when something looks concerning. It never diagnoses or gives medical advice; it only flags and escalates.
- Owner App (Facility Admin / Nurse): an identity and access control console for the agent. Admins can:
- Authenticate each agent with its own identity, scoped to a single patient
- Allow only selected MCP tools (e.g.
read_ehr,read_guideline,alert_nurse) - Watch every tool call in a live activity feed (what data, which tool, allowed or denied)
- Revoke a tool mid-task, with the very next request blocked
- Hit a kill switch to shut the agent down completely
How we built it
Every tool call passes through a per-request authorization check against the permissions table. That's what makes revocation take effect instantly instead of waiting for a token to expire.
Challenges we ran into
- Mid-task revocation: tokens stay valid until they expire, so revoking access didn't stop an agent that was already running. We moved to checking permissions on every single tool call.
- Voice latency: [real-time voice with tool calls in the loop was tricky to keep responsive]
- Graceful degradation: when access is revoked mid-task, the agent shouldn't just crash. We designed it to finish what it safely could, like still alerting the nurse.
Accomplishments that we're proud of
What we learned
- Agents need identities just like people do, and least privilege matters even more because agents drift
- Healthcare AI isn't just a model problem; trust, consent, and auditability are the product
What's next for HumanLoop
- More conditions: diabetes (glucose alerts), COPD (SpO2 monitoring), UTI-related confusion detected from speech
- Real EHR integration via FHIR
- Family/proxy consent views
- Automatic anomaly detection that suggests revocations before an admin has to spot them
