Niyora · My own app
Privacy and safety, built into every layer of Niyora
Women bring Niyora their hardest moments: a fight with a partner, a feeling they can't name, sometimes a crisis. So privacy and safety sit in the architecture, in the AI and in the words. Here is everything Niyora does, and how each promise is built.
Published 6 min read
Checked against the app's code and privacy policy on 18 November 2026, version 3.4.
0
accounts, analytics or ad trackers
AES-256
encryption on her phone
Free
crisis help, always
Never
used to train Google's AI
On this page
Her words stay on her phone
No accounts and no server of mine holding her moments, so there is nothing of hers for me to see or to leak.
- No accounts, no backend
- She never signs up. Her moments are never uploaded to a server I run, so I can't see them.
- Encrypted on the device
- Her words, her responses and the messages she saves are sealed with AES-256 and an HMAC check. The key lives in her iPhone's Keychain and never goes to me or to Google.
- A test keeps it that way
- The build fails if any place that stores her words is left unencrypted.
- No analytics, ads or tracking
- No analytics SDK, no advertising, no third-party tracking and no ad identifier. Another test keeps analytics out of the app.
- Photos aren't kept
- A photo she adds to show Moon what happened is held in memory for that moment, then let go.
- Apple Health is read only
- Niyora reads her period days and never writes anything back. Files she imports from other apps are read on her phone and never uploaded.
- A private lock screen
- Reminders are made on her phone, with no push server, and never name a feeling or the person she wrote about.
- Hers to delete
- She can delete her history in the app at any time, and it keeps only her most recent 300 moments.
What leaves the phone, and what doesn't
Moon needs an AI to reply. Only what she gives Moon goes to it, with identifying details taken out where the code can find them.
- Only what she gives Moon
- Her words, a photo she adds or her voice, and for context a few of her earlier moments on the same topic, go to Google's Gemini through Firebase and Vertex AI.
- Names, emails and numbers swapped out
- Before her words reach the AI, emails, phone numbers and names the code can spot are swapped for stand-ins, then put back in Moon's reply on her phone.
- Her cycle stays private
- The AI learns only whether today is a premenstrual day, never her dates.
- Not used to train AI
- Google doesn't use what she sends to train or improve its models, and no person reads it in normal use.
- Her voice isn't recorded
- A talk with Moon streams to Gemini Live as she speaks. Niyora keeps no audio, and Google may hold a talk for up to 24 hours only so it can reconnect.
- Only the real app can ask
- Moon's text requests are checked with Apple's App Attest, which proves they came from Niyora without identifying her.
- A random ID, nothing more
- When the app checks for updates and settings, it sends only a random installation ID, never anything about her.
- Without AI, nothing leaves
- If the AI is off, Moon uses its own written replies and nothing leaves her phone.
She stays in control
Every step that touches someone else, or anything outside her phone, waits for her.
- She says yes to the AI first
- Before Moon's first reply, a screen says Moon is an AI, names Google's Gemini and asks. Nothing goes to the AI before she agrees.
- She can take it back
- From her You tab, at any time. Nothing more goes to the AI until she agrees again.
- She sends every message herself
- Moon drafts, she edits, and her own share sheet opens. Moon never sends anything to anyone.
- She decides when a talk ends
- Moon never winds a talk down on a count. It ends when she pauses, says goodbye or says stop.
- A report she shares on her terms
- Her report for a therapist is a file she chooses to share. Names are left out and only the relation stays, like her partner or her mum.
Safety comes first
Some moments are more than a bad day. When they are, Moon steps back and puts her safety ahead of everything else.
- Help the moment it's needed
- If her words point to a crisis, Moon stops, shows a calm screen with helplines, and asks and saves nothing more.
- The AI can only add help
- The AI's own safety read can add crisis resources. It can never remove them.
- Crisis help is free, always
- Urgent support, breathing, calm activities, her history and her therapist report are never behind the paywall.
- Abuse changes the whole flow
- If she describes abuse, safety resources come first, and Moon never suggests confronting the person or offers reasons for what they did.
- A safer way to reach help
- The domestic violence line opens as a text first, because a call log is a trace an abuser can find on a shared phone. In a voice talk, Moon ends the talk, since someone nearby could hear.
- Discreet support for eating
- If her words point to an eating disorder, the support screen never mentions food, since it may be read over her shoulder, and Moon never talks about food, weight or bodies.
- A person, when it's heavy
- In the heaviest moments and the biggest decisions, Moon gives no advice and offers a therapist instead. Once she says she already sees someone, Moon doesn't ask again.
- Adults only
- Niyora is for women 18 and over.
What Moon will and won't say
An AI that talks to women at their lowest has to be honest about what it is, and careful with every word.
- It's clear it's an AI
- Before she starts, Niyora tells her Moon is an AI and isn't a doctor or a therapist.
- It never diagnoses or treats
- Niyora doesn't diagnose, treat or give professional mental-health advice.
- No medical advice
- Moon never advises on medication, supplements or hormone therapy, and points her to her doctor.
- It never blames her body
- Moon never puts her feelings down to her cycle, her hormones, ADHD or perimenopause.
- It never excuses harm
- When someone hurt, threatened or controlled her, Moon offers no reason or softer meaning for it.
- It never takes sides
- Moon helps her see a moment from more than one side, and doesn't judge the people in her life.
Built on evidence, checked in code
The AI writes the words. The code decides what's true, what's safe and what ships.
- Facts with their sources
- 156 fact cards, each with its source and evidence grade, each approved by me. The AI can pick one and never rewrites it.
- Sources the AI can't invent
- Every perspective's source label is fixed in code, so the AI can't make up a source or upgrade the evidence.
- Her words are data
- What she types reaches the AI marked as her words, and nothing she types can break out of that mark.
- Every prompt versioned and evaluated
- Every prompt is versioned, and a changed prompt can't merge until a live eval has run on it.
- Safety can't quietly slip
- If the safety tests score worse than before, the build fails.
- Tested on fictional women only
- Every test moment is written from scratch. No real person's words and no third-party transcripts are in the code.
- When the AI fails, help doesn't
- If the AI keeps failing, Moon switches to its own written replies, and the crisis check is never held back.
- The words can't overpromise
- A rule in the codebase: no screen, comment or doc may claim more than the app does.
What I'm still improving
Doing right by her includes saying where it isn't finished yet.
- The name swap has limits
- It catches names with a cue around them, like "my sister Sarah". A bare name can slip through, and photos and voice aren't scrubbed. The privacy policy says so plainly.
- Helplines are US first
- Women elsewhere get a worldwide helpline directory today. Local lines for each country come next.
- Google's abuse logging
- Google may keep messages its systems flag as abuse for up to 90 days. Switching that off needs an enterprise agreement with Google.
Why it matters
Trust is the product. A woman should be able to bring Moon her worst night without wondering who else is reading, and leave safer than she came.
Topics: privacy, safety, trust, AI. Also as plain markdown.