Habitly

The problem

A nutrition and habit product for the Greek market: an onboarding questionnaire turns into a personalised plan with macros, meals and daily habits, with photo-based food logging and an assistant for day-to-day questions.

Make a personal plan feel generated for you, while keeping every number and safety rule out of the model's hands.

The challenges

• ●Nutrition figures carry real consequences, so anything a language model produces has to be treated as a draft rather than a fact.
• ●A plan that fails validation is worse than useless if the fix is to generate a new one — the user loses what they already agreed to.
• ●Photo-based food logging arrives as whatever the phone produces, including formats a browser will not decode on its own.
• ●Several specialists contribute to one plan, so context has to be scoped per stage or the whole history leaks into every call.
• ●Subscriptions, privacy requests and model costs all need to be auditable, which rules out doing them in the client.

The solution

• ●Nine AI specialists behind one product — intake, macros, meal planning, plan finalising, a health advisor, a photo-based food recogniser, a progress tracker and a chat companion — each with a strict input and output schema.
• ●A hard boundary between the model and the truth: AI drafts blueprints and recipes, while the numbers and safety rules come from a deterministic catalogue compiled from a source ledger.
• ●Shared memory derived only from real events in the app; assistant output is never treated as evidence about the user.
• ●Validators that check without mutating, and a repair step that runs before anything is regenerated — so a bad plan is corrected rather than rerolled.
• ●Subscriptions, admin operations, privacy requests and model-usage accounting all run as server-side functions rather than in the client.