AI Therapy Notes Assistant
HIPAA-conscious speech-to-SOAP-note generation, in production with a real clinician.
- ~70%
- Less documentation time
- 2 wks
- To real-world adoption
- 0
- PHI exposure points
A licensed mental-health counselor was spending a chunk of every evening writing up the day's sessions from memory. This tool listens to a full session and produces a structured SOAP note — but the whole engineering challenge was doing that without ever putting protected health information somewhere it shouldn't be.
The problem
Clinical documentation is a tax on care: hours of after-hours typing, and every reconstruction from memory risks losing detail. The obvious AI fix runs straight into the hard part — this is some of the most sensitive data there is.
What I built
- End-to-end pipeline: WhisperX transcription → Azure OpenAI generation → structured SOAP output, fronted by a Next.js app and a FastAPI backend.
- Architected secure API boundaries and data isolation so PHI stayed contained across every processing layer — zero exposure across transcription and generation.
- Tuned the generation stage on full-session audio rather than snippets, so notes reflect the arc of a session, not just keywords.
Outcome
Adopted by a practicing counselor within two weeks of launch, replacing a fully manual workflow and cutting post-session documentation time by ~70% while removing transcription errors.