← Projects

Scribe32: AI Clinical Scribe

Braces4You · Designed, built and operated by me · 2025 – Present

What I did: I built Scribe32 end to end and I'm its sole maintainer. I designed the AWS architecture, wrote the Python backend and the web app, set up the CI/CD and the four environments, and I own incident response and releases. The practice uses it every day.

11daily users: 1 orthodontist, 10 clinical assistants

< $6/mototal AWS + Deepgram spend

~$0.45per user per month, vs $75 for the commercial scribe it replaced

4environments: prod, staging, demo, local

What it does

An assistant picks a patient and a note template, hits record, and has a normal conversation. When the visit ends, Scribe32 transcribes the audio with speaker labels and turns it into a structured clinical note (SOAP or a practice-specific template) that they review and paste into the chart.

Scribe32 New Visit screen with patient search, template picker, Start Recording button, a provider leaderboard and a patient transcripts chart — New Visit: record or upload audio, plus office activity (recordings per provider, transcripts over the last five working days).

Scribe32 visit detail showing a generated SOAP note above the speaker-labelled transcript — A generated note with its diarized transcript.

Scribe32 Visit History list with search, date and provider filters — Visit history, searchable across every visit by patient name.

Screenshots are from a local build of the app running against its mock API, so all names and visits are made-up sample data. No patient data is shown.

Stack, and why

Layer	Choice	Why
Compute	AWS SAM, Lambda (Python 3.12), API Gateway	Usage is bursty: busy clinic hours, idle nights and weekends. Pay-per-request serverless is what keeps the bill under $6 a month, and there are no servers for a one-person team to patch.
Data	DynamoDB (notes, patients, templates, users, audit log)	On-demand pricing, no connection pooling from Lambda, KMS encryption at rest and point-in-time recovery on the tables that matter.
Speech-to-text	Deepgram (Nova-2, diarization, smart formatting)	Accurate on clinical vocabulary, fast, and separates speakers so the note knows who said what. The API key lives in Secrets Manager.
Note generation	Amazon Bedrock (Claude Haiku)	A HIPAA-eligible service that keeps transcripts inside the AWS account, and a small fast model is plenty for structuring a transcript into a template, at a fraction of the cost.
Auth	Amazon Cognito with role-based access	Admins invite staff, manage templates and see the team; assistants just record. The API uses a Cognito authorizer by default.
Frontend	Vanilla JS + Vite	A small, fast single-page app with no framework to keep upgrading. Light and dark themes.
Delivery	GitHub Actions	Checks run `sam build` on every change; deploys go to staging from the staging branch and to prod and demo only from main.

Considerations

HIPAA-aware by design. Per-function least-privilege IAM (the generate function can only invoke Claude models), KMS encryption on every table, CloudTrail, and API access logs with the caller's Cognito identity.
Audit trail that can't break the app. Every account and admin action is dual-written: a structured line to CloudWatch Logs and a row in a DynamoDB audit table with no TTL. If the audit write fails, it's logged and the user's action still succeeds.
Cost guardrails. AWS Budgets alerts, right-sized Lambda memory and timeouts, and a public demo that is rate-limited per IP so it can't run up the bill.
Safe releases with one maintainer. Separate prod, staging and demo stacks plus a local mock API mean I can test a change end to end without touching real patient data.
Built around the clinic's workflow. Custom templates per visit type, one-click copy into the chart, search across every visit, and a leaderboard so the office can see adoption.