Honu
Training, food and habits, linked to the care team
-
-
-
How it works
- Appone-tap log, meal photo, voice
- honu-aireads the photo, transcribes audio, answers
- Supabasedata in São Paulo, private by default
- Consentthe client chooses what the team sees
- Pro dashboardwho went quiet, adherence, draft message
- Suggestionplan and message go back to the app
-
-
-
-
-
-
App screens made for the App Store with sample data; the professional dashboard is in demo mode, with fictional clients.
What I did
- A phone-first app built for short moments: one-tap logging, AI-read meal photos, habits in seasons and body measurements logged on a body map.
- A text or voice assistant that answers from the person’s own logs (“what should I have as a snack?”), with audio transcribed on the server. It doesn’t prescribe diets or medication.
- A table of 1,956 foods (TACO, IBGE and USDA) with the source of every value. Trace amounts and missing data never become zero.
- On the other side, the professional area: who needs attention, per-client record and an invite with a consent form. The client chooses what to share.
- Demo mode with a fictional gym, to present to partners without touching the database.
- The same codebase packaged for iOS and Android, with the account deletion and privacy the stores require.
Stack
- Vanilla JavaScript
- Supabase (São Paulo)
- RLS
- Edge Functions
- Vision AI
- Whisper
- Capacitor 8
- Vercel
- Playwright (Python)
Private code: I walk through the architecture and tests on a call.
Numbers
- 146Playwright test scripts
- 1,956foods, every value sourced
- v138versions shipped
Decision
A message drafted with AI reaches the client labelled “text suggested by Honu and reviewed by” the professional, because of Brazil’s CFN resolution 856/2026. The AI organizes and suggests; a person signs off.
What broke
The regression suite grew past two hours and some tests got stuck on environment dependencies. It now runs in two halves, the 3D test runs on its own, and known failures are logged with their version.