MedScribe AI

Featured Privateai

HIPAA-compliant SaaS that transcribes therapist voice memos into structured SOAP clinical notes using AI. Real-time speech-to-text via Deepgram, AI-driven note structuring with Google Gemini, encrypted patient data storage, and session management.

Next.jsSupabaseGoogle GeminiDeepgramTypeScript

Therapists spend 2 hours/day on notes. This gives them back 90 minutes.

The problem

Mental-health clinicians spend roughly a third of their workday typing SOAP notes after sessions — Subjective, Objective, Assessment, Plan. The notes are formulaic but high-stakes (billing, legal, continuity of care). General-purpose transcription doesn't structure them correctly, and clinicians shouldn't have to wrestle with prompt engineering.

The approach

Built a HIPAA-aware workflow: clinician records a voice memo after a session, Deepgram streams real-time transcription, Gemini structures the output into proper SOAP sections with the clinician's prior notes as context, and the result lands in Supabase as encrypted PHI tied to a session record. Clinician reviews and approves before anything is committed to the chart. Every read/write of PHI is logged for HIPAA audit.

Tech decisions

Supabase

Row-level security + Postgres gives clean per-clinician data isolation; HIPAA-compliant tier covers the BAA

Deepgram for STT

Best-in-class medical vocabulary handling; streaming reduces perceived latency

Google Gemini for SOAP structuring

Long context window holds the prior chart + session transcript together

Clinician-approves-before-commit

AI is a draft generator, not a decision-maker; medical liability stays with the human

Encrypted at rest + in transit

Non-negotiable for PHI; baseline HIPAA technical safeguards

Outcomes

Real-time STT during dictation — no batch wait
Auto-structured SOAP notes ready for clinician review
HIPAA-aware storage with full audit logging
Session-to-session continuity — prior notes feed into next session's context

What I learned

In regulated workflows, the AI feature is a starting draft, not a finished artifact. Designing the human-in-the-loop step well is what makes the product trustworthy. Real-time STT changes the dictation UX more than any model upgrade — perceived latency is the whole game.