Inside Math Wizard: Privacy-First Early Maths with Apple Intelligence

The interesting question about AI in a children's app isn't whether a model can write another maths question. It's whether software can adapt to a child without learning about the child.

A four-year-old benefits when an app notices they're stuck and explains the same idea a different way. That same four-year-old shouldn't have to hand over a name, a voice history or a behaviour profile in exchange.

This is how Math Wizard, an activity in Meet Anvi 1.6, is designed. It's part of Anvi Plus, which a parent turns on behind the PIN. We made the decisions that protect a child at the start, because you can't add them later, and this article explains them.

The whole design rests on one rule:

AI teaches. Code decides.

Maths that feels like play

India's National Curriculum Framework for the Foundational Stage (NCF-FS) puts play at the centre of learning from ages 3 to 8. It describes free, guided and structured play, and names maths games, stories and guided conversation as ways to reach learning goals.

So Math Wizard is built from short missions, not worksheets. A mission is five puzzles, then a natural stop. A child feeds a bunny four carrots, works out which basket has more apples, or finishes a pattern on a train. Each mission has a precise goal; the child just experiences it as a small story.

It builds on what children already do in Meet Anvi: Count goes from 1 to 100, and Blocks and Art sort shapes and colours. Math Wizard adds the next step, connecting quantities to numerals and then reasoning with them.

There are three levels. They're product bands, not judgements about a child: a child can move up or down one skill at a time.

The eight skills are the same at every level and get harder within it: Count & Collect, More or Less, Number Match, Number Order, Patterns, Shapes, Add & Subtract and Story Math.

Where Apple Intelligence fits, and where it doesn't

Apple's Foundation Models framework gives apps a language model that runs on the device, on iPhones and iPads that support Apple Intelligence. Math Wizard uses it only there, and only while the app is in English. In every other case it uses its pre-written hints. Apple's own developer guidance is candid about its limits: the on-device model isn't designed for basic arithmetic or complex reasoning.

That's exactly the boundary a children's maths activity needs. Suppose the activity shows:

3 apples and 2 more apples

Ordinary code generated that problem and already knows the answer is 5. The model never decides that. What it can do is choose how to explain it when a child gets stuck:

"Count the 3 apples, then count on two more."

"Push the new apples next to the others and count them all."

The answer stays fixed. Only the explanation adapts.

How it's built

Meet Anvi is organised in layers that can't quietly reach into each other. The parts of Math Wizard that decide what's true sit in the shared core of the app, the same code on iPhone, iPad and Android. Only the Apple Intelligence part is platform-specific.

Code decidesAI teachesThe child

1. Child taps, drags or choosesWhat the child does on the screen.
2. Math EngineWrites each problem and checks the answer.
3. MasteryA few counts per skill, kept on this device.
4. AI CoachOne short hint, only when the child needs help.
5. Safety FilterChecks the hint. Anything doubtful is thrown away.
6. Hint libraryPre-written hints, for whenever the coach can't help.
7. What the child seesA short, kind hint, or the next puzzle.

The Math Engine owns the answers

Every problem is generated and checked by deterministic code, with its answer worked out before the child sees it. This is what gets tested exhaustively, and the tests run on every platform. Nothing the AI produces can change an answer key.

Mastery is a handful of numbers

For each skill, the app keeps the level, attempts, correct answers, a run of correct answers and when it was last played. No name, no birthday, no transcript. The skill's level moves one step at a time: five right in a row steps up, and mostly misses step down. From the counts it shows one of four plain labels: Practising, Developing, Strong or Mastered. A skill stays on Practising for its first three attempts, and reaching Mastered takes at least six attempts at 90% accuracy.

The AI Coach gets a tiny, structured job

The coach is never asked to "teach this child maths". It's handed a structured description of the moment and nothing more:

• the problem, as numbers and pictures ("three apples and two apples");
• how many tries this is, and the answer the child chose;
• how long it took, and how many mistakes in a row.

It never receives the child's name, age, account, voice or anything they've said. It returns two things: one short hint, and one suggested adjustment. The adjustment is either to keep going, to use more concrete objects, to step back to smaller numbers, or simply to encourage. Whether to act on that suggestion is still up to the app.

The Safety Filter fails closed

Every hint from the model is checked before a child can see it. It's thrown away if it is:

• too short or too long;
• asks for anything personal (a name, a school, where they live, a secret);
• sounds like a friendship ("best friend", "I'll miss you");
• judges the child ("too slow", "wrong", "silly");
• states a number that disagrees with the real answer.

Any rejection, model error or timeout falls back to the hint library.

The hint library means AI is never required

Every skill has pre-written hints for a first try and for later tries, such as "Touch each carrot one by one and count with me." That library is what a device without Apple Intelligence uses, what any device uses when Apple Intelligence is switched off or still downloading, what every language other than English uses (in that language), and what every device uses when the model's answer fails the filter.

The activity is complete without AI. Apple Intelligence only makes the explanations a little more varied where it's available. It never decides who gets to learn.

Privacy, said plainly

Meet Anvi already has a clear shape, and Math Wizard fits inside it rather than inventing a new one.

A parent account, never a child account. A parent creates the account. A child never signs in, types an email or gives any details. Each child profile has a first name, an avatar, a colour and a school level a parent picks; we never ask for a date of birth. A parent switches Math Wizard on and picks its starting level behind the PIN, in Settings > Parent > Controls, and from there it adjusts one skill at a time.

The coach's working data stays on the device. Mastery for Math Wizard is saved as a small file on the device. It isn't synced to the parent's account. That's the same promise Thinking makes: a child's spoken words become text on the device and are never synced.

The AI runs on the device, or not at all. Coaching uses the on-device model only. Apple's Private Cloud Compute is a carefully designed service that doesn't retain data, but it still sends a request off the device, so we've chosen not to use it for anything a child does. When the on-device model isn't available, the app uses the hint library.

No one else is watching. Meet Anvi is in the App Store's Kids Category. It has no ads, no analytics or advertising SDKs, no social features and no links a child can reach. Apple's MetricKit reports for crashes and launch times stay on the device.

A coach, not a companion

UNICEF's guidance on AI and children, and its 2026 report on AI companions, both warn about systems that encourage children to form relationships with software. A maths hint can be warm without being a friend.

Praise describes what the child did, not who they are. There's no persistent character with a memory of the child, and no chat box anywhere: a child can't type to the model at all.

Touch first, voice later

Voice and camera would be delightful: "say the number", "show me five things". They also widen what a child shares. Under COPPA in the US, a recording of a child's voice can count as personal information.

So Math Wizard uses touch only. When voice comes, it'll follow the path Thinking already takes:

• recognition happens on the device;
• the microphone prompt appears only behind the parent PIN;
• no audio is recorded or kept.

Camera activities need a much stronger reason than "it would be fun", and we haven't found one.

Engaging without being sticky

A mission is five puzzles and then it ends, with an easy way to stop. Meet Anvi sends no notifications, and Math Wizard adds no pressure to come back. The UK ICO's Age Appropriate Design Code asks children's services to avoid nudges that work against a child's interests.

The question we care about is did the child understand something?, not how long did they stay?

What parents see

Behind the parent PIN, Math Wizard adds a plain summary:

The summary tells you what to practise together, and a parent can reset it at any time. It doesn't replay what the child did or said, because the app never kept that.

What we check

• Correct maths: every problem type has tests for its expected answer, and they run on iOS and Android.
• The filter against hostile input: the safety filter is tested with cases such as a hint that asks the child to keep a secret from their parents, or one that offers to be their best friend, along with nonsense, very long text, emoji and empty input. Any of these falls back to the hint library.
• Without Apple Intelligence: every puzzle works on Android, on iPhones without Apple Intelligence, and in every language, using the hint library.
• Parents understand it: a parent should be able to say what's stored and what the AI does after a short read. If you can't, tell us.

Apple's own WWDC guidance on prompt design and safety for the on-device model asks developers to handle guardrail violations and evaluate output. A children's app has to go further than that baseline, so we wrote these checks for children.

What's decided, and what isn't

Decided:

• Math Wizard is part of Anvi Plus. A parent turns it on behind the PIN.
• Once a parent has switched it on, each child gets one first go: one mission. After that the tile leaves the child's home screen unless the parent has a plan. A child never sees a price or a lock.
• Anvi Plus isn't sold on Android yet, so Math Wizard isn't offered there yet.

Not decided: voice, and more missions. We'll add voice only the way Thinking does it, on the device and behind the PIN.

The shape doesn't change: deterministic maths, a small and checked AI coach, progress that stays on the device, and an activity that works fully without any AI at all.

Personalise the lesson, not the child.