Nightlight: a Fire TV bedtime story studio. Create a story with the remote, steer the plot, and wind down to sleep.

Inspiration

Bedtime is the one screen-time moment every family shares, yet almost everything on a TV is designed to keep kids watching. We wanted to flip that: a TV app where the child is the storyteller and the story is designed to end in sleep.

What it does

Nightlight is a bedtime story studio for Fire TV (Vega OS). A parent and child build a story together using only the remote:

Story wizard (6 steps, all D-pad): hero (preset or "draw your own"), world, art style, tonight's lesson (bravery, sharing, kindness…), length and narrator.
Illustrated player: slow pan-and-zoom art, word-by-word captions, and choice points, where the child picks what happens next with the remote, with a soft countdown that chooses a default.
Wind-down mode: every page is calmer than the last. The screen warms and dims, the pace slows, and the story ends on a goodnight scene and a night-light sleep screen with a timer.
Library & series: saved heroes and past stories, so "Pip's next adventure" can continue tomorrow.
Parent settings (PIN): allowed lessons, length limits, captions.

How we built it

React Native for Vega (RN 0.83) on Vega SDK 0.24, run on the Vega Virtual Device. 10-foot UI throughout: a 1920×1080 design grid, large focus targets with a clear focus ring, and Back always works.
Amazon Devices Builder Tools (ADBT): the MCP server, Agent Skills and steering docs, version-pinned and committed so every teammate and agent has the same context.
Original content only: illustrations are drawn in code from shapes, so there's no third-party IP.
Reproducible monorepo: pnpm workspace, shared API/event types, pinned toolchain (toolchain.json). From clone to a running app is:

pnpm bootstrap && pnpm tv:sim

CI runs lint + typecheck + 14 tests on Ubuntu and Windows.
AWS backend skeleton (CDK): HTTP API → Lambda (Node 24, arm64), DynamoDB, private S3 + CloudFront. It's ready for the planned all-Amazon AI stack: Nova 2 Lite on Bedrock for story text and drawing-to-hero vision, Polly neural voices with speech marks for synced captions, and Bedrock Guardrails for child-safe content.

Wind-down is planned around a story "energy" value $ e \in [0, 1] $ that falls page by page. Narration speed will follow it:

$$ \text{rate}(e) = 80\% + 20\% \cdot e $$

so the voice slows from 100% at the start to about 80% by the goodnight page.

Challenges we ran into

Retired image models: both Amazon image models (Nova Canvas and Titan Image) reached end of life before the deadline, so we moved to curated original art.
Polly generative voices don't return speech marks, so we chose neural voices to keep word-highlighted captions.
Remote key names: Vega OS 1.2 sends raw names (enter, play) that differ from the documented ones. We measured them on the simulator and normalized them in one place.
No simulator screenshots: we verified the UI with key injection, log reading and a small network "probe".
A broken test mock: the Kepler Jest preset's BackHandler mock is broken, so we wrote our own.
AWS credits didn't arrive in time to switch on live generation, so we shipped the full flow as a polished showcase first.

Every issue is in our friction log, with severity and workaround, and every design choice is in our decision record.

What we learned

Designing for the couch and the D-pad changes everything, from focus states to how long a choice countdown should feel.
Agent tooling (MCP + skills) speeds up platform work a lot, as long as you pin versions.
Check that AI models still exist early. Two of ours were retired mid-hackathon.

What's next

Switching on live AI once AWS credits land:

Story text: Nova 2 Lite with Guardrails.
Narration: Polly, with speed that follows wind-down.
Draw your own hero: the kid draws on paper and a parent snaps it with a phone via a QR code on the TV.
Device testing: on a physical Fire TV.