AlcoTrack

AlcoTrack was the first app I ever built. In 2012, still at school, with my first Android phone, I wanted to build something my friends and I would actually use, and not another to-do list app or a calculator. Fourteen years later it still has users.

And it has the history to show for it, going back to what I started as a teenager in Eclipse. So it kept being rebuilt: most recently in Kotlin, Compose, Hilt and Room, with a migration path so nobody loses ten years of their own history, plus the tooling that should have been there from the start, screenshot tests and automated deployment. Some of the code, the blood alcohol calculation for one, is still at heart what it was back then.

The last thing I built into it is the AI coach, a chat inside the app that answers questions about your own drinking data. Getting it to answer was easy. Getting it to stop overstating things and inventing others was not, and the mistakes were a little sneakier than plain hallucination: it would fetch a real number out of the usage data and then wrap it in a claim no tool had ever checked. Something true and something invented in the same sentence, in the same confident tone. The fix turned out to be in the wiring, not in the prompt. With an agent like this the interesting work is not in the LLM and not in the system prompt, it is in the guardrails that make its confidence earned.

AlcoTrack has been translated into 17 different languages by its own users. By people who just wanted the app in their own language.