An AI copilot that actually put together food, culinary, fitness & emotion in a single platform.

A year ago I started teaching an AI to cook, on evenings and weekends. This week it went live on the App Store and Google Play. Solo build: services, web, native mobile, and the pipelines in between.

CuocoMitra is an AI cooking and nutrition companion. Plenty of apps bolt a chatbot onto a food database. I wanted the inverse: a rigorous system where the AI is the interface, never the authority.

https://www.producthunt.com/products/cuocomitra

These are the design decisions I would defend in any architecture review:

1. The LLM never touches a number.
Models are great at intent, bad at arithmetic. My agents can only emit validated, schema-checked actions (swap this ingredient, adapt to my diet, plan my week). Free-typing an answer is not an option the model has.

2. Deterministic domain engines do the real work.
All food math lives in pure, isolated engine packages: no network access, no storage access, same input gives the same output. Ingredients resolve against a real nutrient dataset through a matching scorer, and macros are computed, never generated. 34g of protein is a calculation, not a hallucination.

3. The AI plane is isolated from the data plane.
Agent sessions, tool orchestration, and live token streaming run in their own service behind their own auth boundary. Model latency or an outage never blocks a normal read or write. Worst case, the product degrades into a great manual cooking app, not a brick.

4. Every AI edit is a versioned mutation.
An adapted recipe is a new version with full history, revertible in one tap. The agent has write access to nothing it cannot undo.

5. Retrieval over memory.
Culinary and nutrition answers are grounded in a curated knowledge base through an embeddings pipeline. Traceable to sources, not improvised from whatever the model memorized in training.

6. One brain across every surface.
Photograph a plate and a vision model recognizes the dish, then the same deterministic engine attaches the macros (I call it MitraVision). Ask for a diet swap and a fusion of culinary knowledge and your learned taste profile keeps the dish authentic: a Thai curry should still taste Thai after the adaptation. One profile of your palate, constraints, and goals drives search, meal planning, and personalization everywhere.

7. Models are replaceable, guarantees are not.
The AI layer is deliberately thin. When a better model ships, it slots in, and every correctness property above stays untouched.

Quietly proud of one more thing: a single typed API contract powers both the web app and the native mobile app, and agent progress streams live to both.

Best engineering classroom I have ever paid for with my own weekends. (Mollie friends: the day job is very safe 😄)

If you build LLM agents, retrieval systems, or you just cook, I would genuinely love your take on the approach.