Why we built an AI coach that never goes online

Focusara is our own focus, habits and wellbeing app. People log their sleep, mood, hydration and habits in it, which is exactly the kind of data that should not be making a round trip through somebody else's cloud.

So when we added an AI coach, we made a rule: it runs on the phone or not at all.

A language model on the device

The coach, Ask Focusara, runs Gemma 3 1B entirely on the device. After a one-time download of about 584 MB, it answers with no server, no API bill and no connection. It works in airplane mode.

Because it runs locally, it can read the user's actual week: the app passes their real numbers in with every question, and the model answers from those.

The hard part was not the model

Small models are eager to please. Ask about your water intake on a week you did not track it, and a model will happily invent a number. For a health app that is worse than no answer at all.

So the coach has a strict rule: if a metric was not tracked, it says so instead of guessing. We also run it at a low temperature, 0.2, so it stays close to the data. Getting that behaviour right took more care than wiring up the model.

What this means for other products

For many features, on-device AI is now good enough. When it is, the cost of each AI request drops to zero and private data never leaves the phone. It is one of the first questions we ask when a client wants to add AI to an app.