Client product · Healthcare AI

Composer AI

A clinical documentation product: a consultation is recorded or uploaded, transcribed, and turned into notes and letters that a clinician reviews.

What I owned

Designer and developer of the product. As backend architect I was the main author of the Django backend: the API, multi-tenant access, the audio and transcription pipeline, note and letter generation, and Stripe billing.

What made it hard

• Consultation audio comes from real clinics, arrives in chunks and is often noisy.
• Every request must stay inside one practice and pass an active-subscription check before any business logic runs.
• Notes have to follow each clinic’s templates and the clinician’s own writing style.

Decision

Clinician review is an explicit state on each consult (pending, reviewed or returned, with reviewer, time and notes), rather than treating generated text as final.

Tradeoff

An extra step for the clinician, in exchange for a clear record of who accepted each note and when.

What shipped

In production on Azure Container Apps across dev, staging and prod.

• Each audio segment is cleaned (noise reduction, high-pass filter, presence EQ, compression, loudness normalisation) to 16 kHz mono.
• Whisper transcription runs through a provider-agnostic client.
• A correction pass fixes medication and clinical-exam spelling without reordering the text.

Open Composer AI (opens in a new tab)