N° 06 / 08 — Mobile · On-device AI
Soie
Your journal, by voice or keyboard, read by an AI that lives in your phone.
- Role
- Design and development
- Year
- 2026
- Status
- LivePersonal
- Stack
- React Native · Expo · TypeScript · llama.rn (on-device LLM) · SQLite · Speech recognition
01 / 07
The problem
A journal holds what is most personal to us. Analysing it with AI should not mean sending it to someone else's servers.
02 / 07
What I did
You tell your day, by voice or keyboard. The next morning, the day before takes on its colour: the AI draws out the main emotion, mood, energy and stress.
An annual “rewind” then lets you go through and analyse your year.
03 / 07
How it's built
A language model is downloaded once, then runs on the phone with llama.rn. Entries are stored in SQLite on the device, and analysis happens in the background.
The app is built with React Native and Expo, for iOS and Android.
- Voice or keyboard
- SQLite on the device
- Local model (llama.rn)
- Recap and rewind
04 / 07
Challenges
- 01
A language model in your pocket
The model weighs several gigabytes: it is downloaded once, then runs on the phone with llama.rn. That means handling the first download, memory, and analysis that happens in the background.
- 02
Reliable analysis, in French
Getting a structured answer out of a small model, always in the right format: the day's emotion, mood, energy and stress.
- 03
Two stores, one codebase
Dictation goes through the phone's speech recognition, and entries are kept in SQLite. The app, in React Native and Expo, is built and published with EAS on the App Store and Google Play.
05 / 07
Results
Soie is available on the App Store and Google Play.
06 / 07
What I learned
A model of several gigabytes on a phone changes the design: you have to handle the first download, background processing and memory. And learn to get reliable structured analysis out of a small model, in French.
07 / 07
Tech stack
- Front end
- TypeScript
- Data
- SQLite
- Mobile
- React NativeExpoNativeWindEAS (builds and releases)
- AI
- llama.rn (on-device LLM)Speech recognition
- DevOps, cloud and testing
- Git
Next project · 07 / 08
Dédale
Dédale
