AiDEPT for Parents

Photograph the homework. Watch the solution come to life.

Comment

Inspiration

Homework time is one of the most common sources of parent-child conflict in households with young children. When a child gets stuck on a math problem, parents default to explaining with numbers and words - and that often escalates into frustration on both sides instead of understanding. We built AiDEPT for Parents so that a parent's role can shift from "frustrated teacher" back to "supportive companion," with the app doing the step-by-step explaining.

What it does

  1. A parent photographs the homework problem (a crop guide keeps the shot focused on the page, not the child's face or home).
  2. Vertex AI Gemini reads the problem and breaks the solution into 2-6 steps, decided dynamically per problem - there are no fixed templates.
  3. Each step is shown as an animation following a concrete -> pictorial -> abstract sequence: quantities appear as dot groups first, then as digits, mirroring the Concrete-Pictorial-Abstract (CPA) approach used in Singapore Math.
  4. Every step includes a checkpoint quiz the child must answer correctly before moving on. This is Human-in-the-loop, not a passive video - kids reach their own "I get it!" moment instead of watching one.
  5. Parents can review their child's solved-problem history at any time.

Sign-in is via the parent's Google or Apple account, and parental consent is required before first use. Photos are sent to Google Cloud (Vertex AI) for analysis and are not stored on our own servers; only structured (non-image) data such as the recognized problem text is retained.

Monetization is transparent: 3 free problems per day, then unlimited access via a monthly (Y980) or annual (Y9,800) subscription - powered entirely by RevenueCat, no ads.

Solving problems earns stamps. The child collects one stamp a day on a monthly calendar, and points toward opening new sticker series - fruit, flowers, animals, sea creatures, dinosaurs and the characters from our own children's programme. Ten points opens one series of twelve stickers, and the child picks which sticker to place.

How we built it

Expo / React Native client (iOS and Android), Vertex AI Gemini for both image recognition and step-by-step solution generation, Firebase (Auth + Firestore) for accounts and data, a Cloud Run backend (asia-northeast1) that mediates every AI call so no API keys live on the client, and RevenueCat for subscription management.

The animation layer is a single generic StepStage component driven entirely by data Gemini returns (targets plus an op like appear/split/merge/rotate/highlight) - there's no fixed template per problem type, so the same engine handles number decomposition, multiplication word problems, and shape problems alike.

Challenges we ran into

  • Getting Gemini to produce a pedagogically correct decomposition (building up from the smallest unit instead of just following the sentence order of the word problem) took several rounds of real-device testing with an actual family and prompt iteration.
  • Apple's review rejected our first submission twice: once for offering only Google sign-in (Guideline 4.8, fixed by adding Sign in with Apple) and once for a missing in-app link to our subscription terms from the paywall (Guideline 3.1.2(c)).
  • Supporting in-app account deletion (Guideline 5.1.1(v)) exposed a real abuse loophole: deleting and recreating an account would silently reset the daily free quota. We moved the free-tier counter from being keyed by account to being keyed by an anonymous device identifier so deleting an account no longer resets it.
  • We also found and fixed a backend signature-verification gap during hardening, and re-verified in production that legitimate requests still succeed after tightening it - a security fix that breaks real traffic is just a different kind of failure.

Accomplishments that we're proud of

The step-by-step "dots before digits" animation is, as far as we can tell, a genuinely different way to present a homework solution compared to a static worked-example or a talking-head explainer video - and it's grounded in an actual pedagogical framework (CPA), not just a stylistic choice. The full pipeline (photo, Gemini recognition, Gemini step generation, animated quiz-gated playback, RevenueCat paywall) runs end-to-end on real devices, not a prototype. The subscription purchase and restore flow has been verified end-to-end on a real device via RevenueCat sandbox testing; we do not yet have real (non-test) purchases to report, since the app was only published during the Shipaton window itself.

What we learned

That "explain the answer" and "help a child understand the answer" are different products. Text-first explanations put the whole cognitive load on reading and language processing; visual, incremental, quiz-gated steps let a young child build understanding at their own pace.

What's next for AiDEPT for Parents

Broadening beyond number composition and shape problems to more of elementary math, and expanding language support beyond Japanese.

Built With

Share this project:

Updates

posted an update —

Behind the build: how we shipped two apps in a month with a team of AI agents

We published a write-up of how this project was built, in Japanese, on note.

AiDEPT for Parents and AiDEPT for Abacus were both built and released on the App Store within about a month. I wrote none of the code myself. Working around a day job, I gave instructions mostly from my phone, and a team of AI agents (Claude Code) did the development. My role was to make decisions and do the final checks.

Part 1 (free) covers how the team grew from a single agent to six, split by role, and what actually became the bottleneck: not how fast the AI writes code, but how fast a human can review and decide.

Read Part 1 (Japanese): https://note.com/aigent_co/n/n2cc7ab9a253b

Log in or sign up for Devpost to join the conversation.

Submission history