Paid team work, live app, Jul 2026 to Present
IZIFIT
Website artwork adapted with English labels.
A fitness app with an AI trainer, food scanning and nutrition coaching.
In short
- An hourly launch engagement, initially agreed for one to two months, then extended.
- The app was announced publicly and its AI trainer launched on 22 July 2026.
- The app is live, with paying subscribers.
- IZIFIT moved its budget to marketing and fundraising, and its original developer carries the app on. I work with them part time.
Engineering within the IZIFIT team.
Within the IZIFIT team, I worked across the mobile app, backend and web portals. I built nutrition macro tracking toward steady daily calorie intake, with food logged by name through AI or by barcode through Open Food Facts. Trainer instructions adapt its jokes, care and motivation to users' interactions, training and nutrition. I developed and checked guardrails against health advice, prompt injection and off-topic requests, and resolved support issues in the app and on Telegram.
The engineering decision.
Give each screen only the context it needs, with navigation chips for additional detail. This keeps the trainer on task and lets users read only what matters on that screen.
The result.
The Nutrition Day follow-up reached an accepted review on synthetic cases and was deployed on 27 September 2026.
After release
What broke, and what changed.
After launch, users in support and our testers reported wrong counts of training days, sets and reps. The low-cost models we used didn't reliably follow counting instructions.
I moved the counting into deterministic code. The app now shows the computed parts directly or gives the model the results to phrase. Tests cover the counting, our testers re-checked the cases that had gone wrong, and the reports stopped.
Evidence and scope
Nutrition Day evaluation
- TypeScript
- React Native
- Bun / Hono
The Nutrition Day evaluation used synthetic inputs, including 16 variety responses and eight factual controls. The final correction received a targeted check, not another complete 24-case run, and the review is not an overall accuracy score.