Case study · Health · Kare4u
A companion on the TV for people living with dementia.
Personalised therapeutic programs, made from a family's own photos, music and voices, playing where the patient already looks.
- TypeScript monorepo
- Six client apps
- Built, pre-deployment
The problem
People living with dementia often spend long hours in front of a TV showing nothing connected to their own life, and carers cannot be in the room all day.
Kare4u turns that screen into a companion. The family sets it up once from a web app, and from then on the patient never has to touch a screen.
How it works
- 01Guardian appNext.js web app for the family
- 02APIEvent driven, on Hono
- 03ProgramsQueued pipelines compose each session
- 04MessagingMQTT to every device
- 05PlayersTV, desktop and Raspberry Pi
At the bedside, the Pi watches for signs of agitation on the device itself and can switch the TV to a calm program.
What I built
- An animated companion that speaks in cloned family voices
- Reminiscence, calm and orientation programs, plus custom schedules
- An onboarding questionnaire and media upload with automatic tagging
- Agitation detection on a Raspberry Pi 5 with a Hailo accelerator, which can start a calm program
- TV pairing by display code, with offline fallback screens
- English and French
- Consent, audit, data export and erasure built in
By the numbers
6client apps from one TypeScript monorepo
58tables in the data model, with consent and audit
40automated tests passing, unit and end to end
Decisions worth a look
The patient never touches a screen.Everything is driven by the guardian's setup and by events, so nothing asks a person with dementia to operate an interface.
Music that meets the patient where they are.The calm program follows the Iso-Principle: it starts at the patient's energy level and slows over 25 to 30 minutes.
Privacy over convenience.Browser voice wake was switched off because it sends audio to a third party; a local wake word option replaces it.
Every outside service sits behind an adapter.Voice, language and video models can be swapped without touching the product code.
Stack and my role
- TypeScript
- pnpm + Turborepo
- Next.js
- Hono
- PostgreSQL + Drizzle
- Redis + BullMQ
- MQTT
- Self-hosted voice cloning
- Whisper
- Raspberry Pi 5 + Hailo
I designed it as architect and product owner and directed AI coding agents through the build. It is built, pre-deployment, and I am open to a partner.