WonderLoop TV

Your child's paper drawing becomes an illustrated, narrated bedtime story on Fire TV, in English or Spanish, that ends by asking them to draw the next chapter.

Comment

Inspiration

Our kids swipe through hundreds of short videos a day: bright, fast, forgotten in a second. Yet give a child a crayon and they don't consume a world, they create one.

We wanted the biggest screen at home to work for that crayon instead of against it: a TV app that rewards imagination, brings children back to paper, quietly teaches a second language, and gives families a calm bedtime ritual to share. Technology is part of their lives, so instead of fighting screens we tried to make one give something back.

What it does

WonderLoop TV turns a child's hand-drawn character into an illustrated, narrated bedtime story on Fire TV.

  1. Draw on paper. No tablet, no stylus: crayons and a sheet of paper.
  2. Scan the QR code on the TV with any phone camera. No app to install, any network. (No phone at hand? Pick one of four sample drawings with the remote.)
  3. Meet the hero. Amazon Bedrock (Nova 2 Lite) understands the drawing (name, species, colours, visible features) and the drawing is redrawn as a storybook character that keeps the child's design: same colours, same floppy ears.
  4. Read together. A 5-page story, written for that child's age, with five 16:9 illustrations. Amazon Polly narrates every page while film-style subtitles light up each spoken word in gold, so early readers follow along. Every story can be created in English or Spanish.
  5. Back to paper. The story ends with a narrated drawing challenge ("Draw a ladder to help Bruno reach the kite!"). Scanning that new drawing creates Chapter 2, where the new friend joins the hero. The loop never ends: every drawing becomes a chapter.

For families: up to three child profiles, medals that celebrate creating (not scrolling), a library of sagas and a corkboard of drawn friends, an arithmetic parental gate for grown-up settings, no ads, no tracking, no database: the drawing is deleted from server memory as soon as the TV picks it up.

How we built it

Fire TV app (Kotlin Multiplatform + Compose): a 10-foot UI designed for the remote from scratch: D-pad focus with golden halos, rails with Leanback-style focus restoration, 48 dp overscan safe zones, Compose for TV (androidx.tv Surface) for the reader and home, an exit dialog that follows Amazon's Controller Behavior Guidelines, and a cinema-mode reader that behaves like a Fire TV video player (controls fade out, OK pauses, ◀ ▶ turn pages).

QR bridge: the TV creates a short-lived session (15-minute TTL) and polls it; the phone opens a bilingual web page served by our backend that downsizes the photo to 1024 px in a canvas and uploads it. No account, no companion app.

AI pipeline on AWS:

  • Amazon Bedrock · Nova 2 Lite (Converse API with forced tool use and full JSON schemas) analyses the drawing and writes the story.
  • A story quality engine plans every chapter from a catalog of concrete worlds, problems and drawable challenge objects, follows a 5-step story spine, and runs an automatic quality review with one guided rewrite (clichés, length per age, language leakage, drawable challenge).
  • Amazon Polly (neural voices Lucia / Ruth) narrates every page and the challenge; word speech marks are aligned to the on-screen text to drive the gold word-by-word subtitles.
  • Illustrations use Gemini image generation behind an IImageAdapter interface (there is no GA Amazon image model with reference-image consistency for new Bedrock accounts yet); every AWS call also falls back to Gemini automatically.
  • Vision → portrait run in sequence to keep the character's identity; then the five illustrations and the narration run in parallel. A full story takes about 12 s on our EC2 backend (vision 1.8 s, story text 5.2 s, five illustrations 5.3 s).

Infrastructure: a Node.js/TypeScript backend in Docker on Amazon EC2 (t3.small, Elastic IP) behind Caddy with automatic HTTPS. AWS IAM instance role with a least-privilege policy (bedrock:InvokeModel + polly:SynthesizeSpeech), so no AWS keys live on the server. App key header, rate limits, image pixel caps and no persistent storage.

What we built during the hackathon

WonderLoop started as a small Gemini-only mobile prototype started on September (phone camera → story). During the submission period we built everything that makes it a Fire TV product:

  • The entire Fire TV app: 10-foot UI, remote navigation, Compose for TV, profiles, library, friends, parental gate, exit dialog, onboarding, launcher icon and banner.
  • The phone-to-TV QR bridge (sessions, mobile upload page, live status on the TV).
  • The move of vision, story and narration to Amazon Bedrock (Nova 2 Lite) and Amazon Polly, with automatic Gemini fallback.
  • The story quality engine (catalog, story spine, automatic review and rewrite) and sagas with narrated drawing challenges.
  • The cinema reader with Polly speech-mark subtitles and the English/Spanish experience.
  • The AWS deployment (EC2, Caddy HTTPS, IAM role), security hardening and a detailed friction log.

Challenges we ran into

  • TVs have no camera. We designed a phone-to-TV bridge that needs no app and works on any network.
  • Text vs. artwork on a TV. A reading panel at 10-foot sizes covered 55% of the illustration. We switched to film-style subtitles synced to Polly speech marks.
  • Reliable JSON from Nova 2 Lite. Plain-text JSON broke in English (typographic quotes, fences); a permissive tool schema returned empty objects. Forced tool use with complete schemas fixed it (6/6 successful calls).
  • No GA Amazon image model for a new account (Nova Canvas reached end of life), so illustrations stay on Gemini behind an interface, ready to swap.
  • Compose for TV is Android-only while our UI lives in Kotlin Multiplatform commonMain; we bridged it with thin expect/actual primitives.
  • Focus on Fire TV: focus lost after dialogs and escaping the end of rails, found only on the real device and fixed with focus memory and key guards.

Accomplishments that we're proud of

  • A child goes from a crayon drawing to a narrated, illustrated story on the family TV with only seconds of waiting (about 12 s to generate a full story), without installing anything on a phone.
  • Stories that read like real picture books: in our evaluation of 16 stories, concrete and drawable challenges went from 38% to 100%, and pages within the age word target from 50% to 100%.
  • Word-by-word gold subtitles driven by Amazon Polly speech marks: the TV reads with the child, not for them.
  • Privacy by design: no database, no ads, no analytics; the drawing is deleted from memory once the TV receives it.

What we learned

  • A TV app is not a big phone app: focus, distance, overscan and the remote change every decision.
  • Borrowing Fire TV's own video-player conventions (subtitles, auto-hiding controls, OK to pause) made a reading app feel native.
  • Structured output from LLMs needs a contract (forced tool use + schema), not just a polite prompt.
  • Small details matter to children: a hero that keeps their colours, and a happy face even when they drew a straight mouth.

What's next for WonderLoop TV

  • Fire TV Appstore launch, with a free tier and an optional subscription (the billing integration is already in place, disabled in this build).
  • An Amazon image model on Bedrock as soon as one with reference-image consistency is generally available, to run 100% on AWS.
  • More languages for families raising multilingual children.
  • Alexa voice: "Alexa, continue my story" and voice answers to the story's questions.
  • Printable colouring pages of each chapter, so the story goes back to paper once more.

Built With

Share this project:

Updates

Submission history