W05 Products · Fitness and nutrition
Kai
A voice-first gym and calorie coach: say the set or snap the plate, and it is logged.
S01 Snapshots
Kai, as it ships.
Real captures of the build: the desktop view, the phone, and the whole page.
C01 What it is
The build, in passes.
You talk to Kai ("bench 80 for 8, felt heavy", "adobo with two cups of rice") or photograph a plate, and Kai logs it. Calories burned come from the watch through Health Connect. The mobile app is Expo and React Native, Android first; the marketing site is Next.js and shares the app's visual pieces.
- Voice loop: speech to text, a fast on-device path for set grammar, AI for the rest, then speech back
- kai-agent edge function: utterance to OpenAI tool calls to the database to a spoken reply, with undo
- kai-meal: plate photo to a vision model with structured output to a logged meal
- Live talk relay that holds the provider key server side, with quotas and transcripts
- Unit tests plus AI evals for voice intents and meal photos
C02 How it is built
Wired from the real dependencies.
Every node is a package the repository actually depends on. Hover one to light its neighbours.
- Expo
- React Native
- Next.js
- TypeScript
- Supabase
- TanStack Query
- Three.js
- Lenis
C03 AI in it
The signal chain.
Turns a spoken set or a plate photo into a logged entry, with a spoken reply.
- 01Voice
- 02Speech to text
- 03Set grammar or model
- 04Tool calls
- 05Database
- 06Spoken reply
C04 Evidence
What the repository says.
Committed under my name and my VibeBox studio accounts. The frame above is drawn in code from this page’s own data; it is an illustration, not a screenshot.