Niyora · My own app
Designing Moon: I tested the AI first
Private by default meant trying a model that runs on the phone. So in late July I fine-tuned one and tested it before deciding.
DraftPublished Updated 1 min read
This is a working draft. I’m still filling it in.
On this page
I tried AI on the phone first
I fine-tuned a small open model on a few thousand practice examples and ran it on my own iPhone. It was good at reflecting feelings back, and weak at suggestions.
It finally ran on my iPhone on July 27, after a full day of debugging. The model itself was fine. Every fix was in how the app loaded it.
Why the on-device model didn't ship.
What the test shaped
That finding shaped the whole design. Moon listens and reflects first, and its suggestions stay grounded in the research.
A reflection she can’t act on is just a nice sentence. Suggestions are the half that has to work.
- For version one I moved Moon to Gemini through Firebase. Moon’s home screen shipped on July 31.
- Small model, funny surprise: the word “moon” in a prompt made it write poetry. Now a test fails if that ever sneaks back in.
Moon never pretends to be human. It never calls itself a person, and there’s no endless chat to keep her typing.
Reflect, Regulate, Respond
Moon went live on August 20 in three steps: think about what happened, calm down, then choose what to say.

Reflect

Regulate

Respond
What I learned
- Test the model before you design around it. Its weak spot told me what Moon should be.
Topics: AI, product design, women's health. Also as plain markdown.