Refuel AI · Platform
AI with context. Behind a safety layer.
Refuel AI is not a chat window bolted onto a tracker. It is a support layer that already knows the week, and knows what it is not allowed to know.
Why AI needs a safety layer in front of it
A general chatbot does not know it is talking to someone eleven days postpartum on four hours of sleep. It answers confidently, sometimes clinically, and it cannot tell the difference between a bad night and a crisis.
Refuel AI is built the other way round. Deterministic safety rules read the moment first. The model receives only a minimised, permitted context. And when a moment needs real people, the app routes to them, and no model can change that decision.
Her context
Stage, recent check-ins, the current focus, whether help was asked for. Minimised and structured, never the whole picture.
Permission filter
Only what she allows leaves the device. The journal never does. Crisis text is handled on the device and never reaches a model.
Safety engine
Deterministic rules read the moment first. If it needs more care, the app routes to real people and real resources. No model can overrule that decision.
Support, not diagnosis
Answers in her language, inside that frame. Never a doctor. Never a prescription. Never an emergency service.
Real screens from the iOS app.
How it works
What happens when she asks something
- 01The safety engine reads first
Fixed rules, on the device, in English and Ukrainian. Crisis language is recognised here and never sent to a model.
- 02The permission filter decides what leaves
Stage, recent check-ins, the current focus, whether help was asked for. Structured and minimised. The journal never leaves the phone.
- 03Our proxy talks to the model
Requests go through Mommy Refuel's own server. The model provider never sees an identity, only the context for that one reply.
- 04The reply comes back inside a frame
Support in her language. Never a diagnosis, never a prescription, never an emergency service.
- 05Escalation stays human
If the rules say a moment needs more care, the app shows real routes to real people. The model has no vote.
Facts
- Sign-in
- Live Refuel AI is the one feature that asks for Sign in with Apple. A private relay email works.
- What is sent
- Only when she uses it: a minimised summary plus the message she typed, to generate one reply.
- What is never sent
- Journal text, private notes, cycle or pregnancy details beyond the stage, crisis text.
- Model
- A third-party model behind our own proxy, swappable without changing the safety layer.
- Analytics
- AI usage is counted. AI content, moods, scores and symptoms never enter analytics.
- Clinical review
- The safety rules are versioned and documented for clinical review. They are a draft pending sign-off before public release.
What Refuel AI is not
- Not a doctor, a therapist or an emergency service.
- Not a chat window bolted onto a tracker.
- Not able to overrule the safety engine, ever.
- Not trained on her data. Her context is used for one reply and not for training.
Questions
What leaves my device?
Only when you use Refuel AI: a minimised summary and the message you type, sent to our server and a model provider to generate a reply. Journal text and private notes are never sent.
Which model?
A third-party model behind our own proxy. The provider never sees an identity, only the minimised context for that one reply.
Can it override the safety layer?
No. The safety engine runs first and has the final say.