- AI / LLM
- MongoDB
- Python
- React
- Typescript
InnoviScribe: AI Medical Scribe
An AI medical scribe that listens to a patient visit and turns it into a structured clinical note, with ICD-10 and CPT codes verified against Find-A-Code.
Overview
InnoviScribe gives clinicians time back from documentation. A provider starts a visit in the web, desktop, or mobile app and talks with the patient as usual. InnoviScribe transcribes the conversation live, then generates a structured note (SOAP, H&P, or a custom template) in seconds, complete with diagnosis and procedure codes that have been checked against Find-A-Code’s code data.
Clinicians review and edit the note, regenerate any section, translate it, and export it to PDF or Word, or paste it straight into their EHR. It’s available on the web, iOS, and Android.
My role
I was a core engineer on InnoviScribe and the primary contributor to the FastAPI backend, the Next.js web app, and the admin portal. I built the subscription billing system, hardened live transcription, added offline recording, designed the note-generation pipeline with code validation, and built the trial onboarding flow. The desktop companion and mobile apps were team efforts.
Highlights
- Resilient live transcription. Audio streams over WebSockets to Deepgram with speaker separation; automatic reconnects, keep-alives, and a circuit breaker keep visits recording through flaky connections.
- Offline-first recording. If the network drops, the browser keeps recording locally in IndexedDB and uploads the audio in order once it reconnects, so a visit is never lost.
- Parallel, streaming note generation. Each template section is generated concurrently and streamed to the screen as it is written, with per-section regeneration and one-click cancel.
- Verified medical codes. The AI must look up every ICD-10 and CPT code through Find-A-Code tools before adding it to the note, and only active codes that match the documented diagnosis make it in.
- A multi-provider LLM layer with quality tiers. Each template picks a tier that balances speed, cost, and reasoning depth, and weak or failed sections are automatically retried on another model.
- Consistent clinical output. Section-level validation and cleanup keep notes in a predictable structure and grounded in what was actually said during the visit.
- Complete SaaS billing. Subscriptions, pro-rated upgrades, auto-renewal, invoices, and promo codes, plus an admin portal for accounts, usage, templates, and trials.
- EHR workflow. A desktop companion types finished notes into Epic using its SmartList syntax.
Tech
Python, FastAPI, WebSockets, MongoDB, Next.js, React, TypeScript, Redux Toolkit, Tailwind CSS, a rich-text note editor, a Vite + React admin portal, Docker, Deepgram, and a multi-provider LLM layer.