AI scribe
Transcribes an uploaded consultation, labels doctor and patient, and drafts a SOAP note with the key entities pulled out.
Healthcare · AIProduct build · 2026
An EHR that writes the notes.
Hospital records, with an AI scribe in the room.
Medicas is a multi-tenant hospital EHR. I architected it and wrote the whole backend and AI service, and built the web app with one frontend developer.
- ROLE
- Architect and lead engineer
- WHEN
- Mar 2026 – Aug 2026
- FOR
- Hospitals, each on its own subdomain
- STATUS
- In development
THE HEADLINE
Talk → SOAP note
Record the consultation, get a draft note the doctor reviews and signs
- 680+
- GraphQL operations
- 1 DB
- per hospital
- Local
- speech and language models
01 · The problem
Doctors spend the visit typing.
In a busy hospital, notes get written during the consultation or long after it. Medicas lets the doctor talk to the patient, then hands them a structured draft to check, so the record is complete without the keyboard in the way.
02 · What I built
A whole hospital, plus the scribe.
Clinical records
Patients, encounters, vitals, labs and imaging, immunisation, referrals and medication administration.
Hospital operations
Appointments, waiting room, wards and beds, maternity, surgery, pharmacy and rosters.
Billing and insurance
Bills, payments and insurance claims in the same record as the care.
Telemedicine and chat
Video visits, real-time chat and push notifications for staff.
Super-admin
Onboard a hospital, give it a subdomain and its own database, and manage roles and permissions.
03 · Key decisions
Accuracy first, then speed.
Tested the model on our own accent
I compared a Nigerian-accent speech model with a general one on a simulated consultation. The accent model invented drug names and dropped the medication advice, so I chose the general model, at 20% more latency.
A database per hospital
Each hospital’s data lives in its own database, routed per request, with the connection details encrypted. One tenant can never read another.
Drafts, never decisions
The AI only writes drafts. Nothing enters the record until the doctor has reviewed and signed it.
Models that stay in the building
Speech-to-text and summaries run on local models, so patient audio doesn’t have to leave the hospital’s infrastructure.
04 · Stack
Under the hood.
Backend
- NestJS 11
- GraphQL
- TypeORM
- PostgreSQL
- graphql-ws
Frontend
- Next.js 16
- React 19
- Chakra UI v3
- Apollo Client
- TipTap
AI service
- Python
- FastAPI
- Whisper
- pyannote
- Ollama
05 · Results
What it added up to.
- 100%
- Of the backend and AI service written by me
- ~30
- Clinical and operational modules
- 5 months
- From first commit to a full hospital system
Next case study · 05
A hospital system, scaled 30×
Bringing AI into clinical work?
I build healthcare software that keeps the doctor in charge.