All-med

Inspiration

In healthcare, missed appointments, delayed billing, and miscommunications cost clinics thousands of dollars and negatively impact patient outcomes. However, fully autonomous AI agents are often too risky for healthcare due to strict HIPAA regulations and the potential for hallucinating Protected Health Information (PHI). We were inspired to build a system that bridges the gap: bringing the power of CALL-E's voice agents to medical clinics while strictly enforcing "Safety by Default" through Human-in-the-Loop (HITL) consent gating and EHR integration.

What it does

All Med is a centralized operations dashboard that listens to Electronic Health Record (EHR) triggers (like OpenDental or FHIR APIs) and automatically generates outbound CALL-E call plans for issues like missed appointments or billing errors.

Crucially, it does not dial automatically. It generates a proposed script, scrubs it for excessive PHI, and places it into a "Pending Consent" queue. A human administrator must review the script, ensure compliance, and explicitly click "Approve" before the system dispatches the call to the CALL-E API. The dashboard then polls the CALL-E backend and tracks the live status of the call until completion.

How we built it

Backend: Python and FastAPI.
Integrations: OpenDental API adapters to ingest clinic events.
Agent Skills: We built custom .md Agent Skills for specific workflows (Appointments, Billing, Lending).
Voice Orchestration: We deeply integrated the CALL-E API. When a human approves a call, the server securely POSTs the call task to CALL-E, maps the required metadata, and continuously polls the v1/calls endpoint to update the UI with live statuses.
Frontend: Vanilla JavaScript and CSS with a focus on modern, responsive, glassmorphism design.

Accomplishments that we're proud of

We are incredibly proud of our Human-in-the-Loop (HITL) Consent Gate. By preventing any automated system from making a phone call without explicit human approval, we perfectly aligned with CALL-E's core philosophy of Safety by Default. We proved that voice AI can be safely deployed in high-risk environments like healthcare if the right architectural safety boundaries are respected.

What we learned

We learned the intricacies of the CALL-E API, specifically regarding outbound Caller IDs and payload structures. We discovered that keeping provider separation clean—where the dashboard handles the scheduling/consent and CALL-E solely handles the execution—makes the architecture incredibly resilient and portable.

What's next for All-med

We plan to expand our adapters to support standard FHIR endpoints (Epic/Cerner) and implement bidirectional webhook listeners so that CALL-E can proactively push call transcripts and structured results back into the patient's EHR chart automatically.