Hewie — AI Health App

Hewie wanted to build a health app that takes people's questions seriously without pretending to be a doctor. The product has to be useful — actually help people understand symptoms, suggest what to track, point them toward the right kind of clinician — while being relentlessly clear about what it is and isn't. Health AI is a category where bad design literally hurts people.

The technical challenge wrapped around an architectural one: how do you build conversational health guidance where the safety layer is structurally enforceable rather than just promised? A chatbot wrapped around a general-purpose LLM, however carefully prompted, is one clever input away from giving advice it shouldn't. That posture isn't defensible in a health context.

We separated the system into two services and treated the boundary between them as the safety guarantee.

The mobile app — React Native, shipping to both iOS and Android from a single codebase — handles the user experience. It never talks to the LLM directly. Every conversation turn goes through a backend API we built in Python, which sits between the app and the model service.

The model itself runs on its own dedicated service, isolated from the consumer app, the mobile build, and the user database. The Python API layer is where the actual product logic lives: structured intake patterns for symptom characterization, deterministic safety rules that fire on specific symptom combinations, escalation messaging that overrides anything the model might say, and rate-limiting per user. The LLM is one component of a structured system, not the entire system.

This architecture is the product differentiation. The safety logic is written in code, not in the model's prompt. It's auditable, version-controlled, and impossible for a clever input to bypass. Specific symptom patterns trigger automatic clinician-referral messaging — the app stops giving advice and explicitly says "this is a thing to see a doctor about, here's why." Other patterns trigger "track this for X days and check back" rather than premature reassurance.

Hewie shipped with a posture that's defensible: useful for symptom characterization and education, explicit about its limits, and engineered so the safety layer doesn't depend on the model being well-behaved. The model service can be swapped — different provider, different model, different version — without touching the safety rules, because the safety rules live in the API layer that sits in front of it.

What we wouldn't do, and didn't do: ship an unguarded chatbot wrapper around a general-purpose LLM. Health is a category where that approach is irresponsible — and the engineering investment in proper architectural separation is the actual product, not a nice-to-have.