Jess Drawing World

Teaching a young child to draw is a real, specific problem: kids learn by tracing and by following small steps, then need encouragement to keep going. Off-the-shelf kids' apps tend to solve this with ads, accounts, leaderboards, and cloud sync — exactly the things a parent does not want pointed at their child. The brief was to build something genuinely useful for one real kid, Jess, that a parent could hand over without worrying about who is collecting what.

That meant taking privacy as a hard constraint, not a marketing line. No account to create, no ads to serve, no analytics backend phoning home, and nothing uploaded to anyone's cloud. But the features parents actually want — turning a photo into a line drawing, a 'how close did I get?' accuracy score, read-aloud for pre-readers, daily reminders — are usually the exact features that lean on paid cloud APIs. Doing them with zero cloud and zero per-call cost is a much harder engineering problem.

On top of that, the most engaging feature — anchoring a reference image onto a real sheet of paper through the camera so a child can trace it by hand — is markerless AR. There is no off-the-shelf Flutter plugin that does this well, and ARCore is a native, low-level surface. The build had to reach down into Kotlin, raw ARCore and GLES2, and still come back up into a single, tidy Flutter app that ships to a real Pixel device — without a sprawling budget or timeline.

We built it as one Flutter codebase (Dart, Riverpod 2.x code-gen state, go_router navigation) targeting Android first, with iOS configured and a Web build behind it. That single codebase carries 250 step-by-step lessons (~1,000 steps) grouped into themed tracks and a visual skill tree with unlock gates, ~285 original IP-clean tracing templates with search and recents (plus ~1,040 inspired-by templates a parent can hide with one Original Edition toggle), and a clean lesson-to-tracing handoff — the everyday loop a child actually uses. Accessibility was built in, not bolted on: 48dp tap targets, Semantics everywhere, system text scale, and a pre-reader long-press tooltip that pairs an icon with a word and speaks it aloud. The app ships in 8 languages with an in-app picker, and a CI i18n guard fails the build if a new hard-coded string slips in.

For the hard parts we went native and on-device. The markerless AR is a hand-written Kotlin PlatformView — ArView, BackgroundRenderer, QuadRenderer, GlUtil — driving raw ARCore and GLES2 to lock a reference image onto real paper, with flip, grid density, pinch-about-point and a Lock/Freeze control. The photo-to-line-art and Pop art filters use a custom adaptive-threshold computer-vision routine written in pure Dart (deliberately not raw Sobel, for cleaner child-friendly outlines). Colour-in is a real flood-fill with an alpha outline map so paint stops at the lines, plus undo history. The self-check accuracy score is local CV. None of it calls out to a server.

Storage is local-first and encrypted at rest: Drift over SQLCipher, with the database key held in the OS Keychain/Keystore via flutter_secure_storage and data namespaced per child. Read-aloud uses on-device TTS; daily reminders use local notifications with UTC-instant scheduling and no push server. Quality was held with 33 test files (~103 test cases), golden/visual-regression tests via Alchemist with a pinned font for deterministic renders, and GitHub Actions CI running analyze plus test on a pinned Flutter 3.44.2. The whole thing was built AI-accelerated with Claude Code, which is how a solo-scale effort produced a broad, governed app at speed.

The shipped scope is broad and real. A child can follow 250 guided lessons along a visual skill tree, pick from ~285 original tracing templates, trace against real paper through markerless ARCore AR on Android, use camera-overlay tracing with gyro de-shake and Easy/Medium/Hard guide layers, turn a photo into line art or a Pop art portrait on-device, colour it in with a true flood-fill, make live kaleidoscope art in Mirror Doodle, get a local accuracy score with a ‘where you went off’ heatmap, earn XP/badges/streaks, hear each step read aloud, and export a time-lapse GIF, a printable A4 colouring page, or a formal achievement certificate. It runs in 8 languages with an in-app picker and a choose-your-accent colour. Parents get a two-week activity dashboard, multiple child profiles, a My Artwork gallery, and a PIN lock on settings.

It runs on a real Pixel Fold today as a 119 MB arm64 APK (version 0.1.0+1), built from a single codebase that also produces iOS (configured and building, not yet shipped to TestFlight) and Web targets — though the headline camera, AR and photo-capture features are mobile-only in practice. Crucially, the running cost is nothing: there is no cloud, no API bill, no analytics service, no per-user infrastructure. Agency builds of comparable scope — native AR, on-device CV, encrypted storage, parent tooling, eight locales — are commonly quoted in the tens of thousands of pounds over several months; this was delivered AI-accelerated by a solo effort, faster and at a small fraction of that. (Cost and time figures are illustrative UK market ranges, not project quotes.)

What it proves is the point. CIA can take a hard brief — privacy as a hard constraint, native AR, on-device computer vision, no paid APIs — and ship a polished, governed, tested app that a parent can actually trust with their kid. Built AI-accelerated with Claude Code, quality-first and cost-disciplined, with the honest caveats noted: it is a pre-release build (0.1.0+1) not yet on the app stores (current distribution is sideloaded APKs), the anchored AR is Android-only today, and iOS is configured but not yet shipped to TestFlight. The capability — and the way we work — is exactly what we bring to client builds.