An AI coach for a US health app

The problem

The client is a subscription app for people losing weight in the United States. The founders wanted the coach to be the product: not a chatbot bolted onto a tracker, but the thing you open, talk to, and trust with your numbers. A coach that forgets you, invents your calorie count or changes your plan without asking is worse than no coach.

The first version proved the idea: members wanted a coach they could talk to. Version 2 was about turning that into a product that could grow, with a coach that remembers each member, acts on their data safely and can be measured before every change. I was brought in to lead it.

What I built

Version 2, from the ground up: the Python backend, the agent, the app on both stores and the native health integrations.

• The agent loop: a streaming Claude agent that decides which of 30+ tools to call to read and update food, weight, sleep and workout data, with every proposed change confirmed by the member in one tap.
• Numbers are computed in code, never by the model. Calories, macros and trends come from functions; the coach explains them.
• Memory: facts extracted from past conversations, a coaching profile that updates daily, and retrieval over a health-document library, so the coach sounds like it knows you because it does.
• Evals that decide what ships: golden datasets plus a blind LLM-as-judge test of coaching quality. They stopped a switch to a cheaper model that scored worse.
• Cost controls: simple requests routed to a small model, prompt caching, and failover between two providers.
• Native work in Swift and Kotlin: Apple Health and Health Connect sync, a camera-based vital-signs scan, and food logging by photo, voice and barcode.
• The revenue path: subscriptions and free trials on StoreKit and Play Billing, and the store releases on both platforms.

Decisions that mattered

• Human confirmation on every write. It costs one tap and removes the whole category of "the AI changed my plan" complaints.
• Measure before switching models. The eval harness exists because a cheaper model looked fine in a demo and was worse in a blind test.
• Two providers from day one. An outage on one is a slow afternoon, not a dead product.

How the team works

I run sprint planning and code reviews for four engineers and work directly with the US founders on the roadmap. AI coding agents are part of the team's daily workflow, with the same review bar as any other change.