Genius Care
A clinical AI agent platform spanning voice reception, clinical documentation, encrypted workflow APIs, EHR automation, and safe human escalation.
Overview
Genius Care is a clinical AI agent platform I led and co-built across voice reception, clinical documentation, encrypted workflow APIs, and a React practice dashboard. Four services run in deployed infrastructure — a 24/7 voice worker, a synthetic canary that dials the agent every fifteen minutes, a scribe service, and an edge API over an encrypted datastore — while the first clinical pilot onboards onto athenahealth. The voice receptionist is the shipped wedge inside a broader platform for durable clinical and administrative workflows, not the definition of the product.
Constraint
Clinical teams lose time and revenue to fragmented front-desk, documentation, scheduling, and EHR workflows. Automating that work safely requires more than a conversational demo: protected data, auditable actions, reliable writes, integration fallbacks, evaluation gates, and clear escalation to people.
Contribution
Built LiveKit voice reception and clinical-scribe workflows over an encrypted edge API: field-level AES-256-GCM whose ciphertext is bound to its own row so records cannot be transplanted, an append-only access log that records column names and row ids but never plaintext, idempotency enforced at the API boundary, and rollback paths. Emergency escalation runs on deterministic tripwires that fire before the model round-trip and lock out every other tool for the rest of the call. The athenahealth integration is built through a shadow-to-failsafe-to-primary ladder, staged behind a flag ahead of pilot cut-over. Python and TypeScript services deploy through Fly.io, Cloudflare, and Vercel behind nine CI workflows and a nightly agent evaluation.
Architecture
Input
System
Outcome
Walkthrough
01 — New encounter
A note opens against a provider, with the patient left optional — a walk-in does not need a record to exist first.
02 — Audio staged
Consultation audio is attached and queued for transcription. Recording in the app and uploading a file both land here.
03 — Structured note
The transcript is drafted into separate clinical sections with an ICD-10 code on the assessment. A section the encounter never covered is left explicitly empty rather than filled in.
04 — Signed
Signing moves the note to the signed list with its provider, encounter time, and code — and clears it out of the drafts queue.
Capabilities
Stack
Next
Want something like it built for your team?
Start a conversation→