Inspiration

In ninety minutes of football, a lot happens. Coaches scribble notes on the touchline, say crucial things at half-time and clarify them in the debrief, but most of it gets lost: notes stay on paper, half-time talks are never written down, and debriefs fade by Monday. The player usually gets a clip and a one-line note. The real intent is scattered, and it often seems to contradict itself. AI can summarise all of that, but a coach can't send a player AI-written feedback that might be wrong. We wanted feedback that is specific, linked to its evidence, approved by the coach, and easy for a player to take in on the screen they already relax in front of.

What it does

CoachOS has two connected apps.

Coach Studio (desktop)

  • Shows the pitch beside one timestamped feed: every statement about a moment, on the match clock, with its source and who can see it. In the demo: a half-time comment (staff only), a note at 75:10, and a post-match clarification.
  • Proposes the principle that reconciles them: attack the channel when the defensive midfielder provides cover. It stays a proposal until the coach confirms it.
  • Ends the feed with a draft message to the player, each section tagged by what it rests on (Match, Coach or AI), and a preview of how it will look on the TV. The coach can edit everything and draw arrows, zones and labels on a tactical frame.
  • Keeps Approve and Publish separate. Publishing is refused while the feedback cites an unconfirmed principle.

Fire TV app (Vega OS)

  • A Who's watching? profile picker, so a shared TV keeps each player's feedback private.
  • A bold streaming-style welcome, then rows for Continue watching, the latest match and past matches, shown as fixtures with club crests.
  • A review screen with a tactical replay of the player's run, the expected run, and the coach's own drawings, plus the coach's instruction up front.
  • A simulated Alexa+ style voice experience. The player asks "Why did that run matter?" and sees the question transcribed. The answer appears word by word with the sources it's based on, and can jump to the moment he got it right. Questions about staff-only conversations are declined.

When the coach presses Publish, the review appears on the player's TV with a New badge within about three seconds.

What's real, simulated or demo data

  • Real: Fire TV app on the Vega Virtual Device; Coach Studio; the publication API; the approve and publish workflow; live sync; the access boundary.
  • Simulated: Alexa (no microphone, speech or Alexa service; questions are picked with the remote). The AI draft and the proposed principle are pre-generated for the demo, with no live model call.
  • Not used: Bee and AWS. Bee capture is planned, not built.
  • Demo data: fictional clubs, players, matches and coach statements. Player photos are AI-generated, and the replay is drawn from position data, not footage.

How we built it

  • Fire TV: React Native for Vega (SDK 0.24, RN 0.83), tested on the Vega Virtual Device. Navigation is D-pad first, using TVFocusGuideView, useKeplerBackHandler for back navigation, and scaling from 1920-pixel designs to Vega's 960 dp layout. Layout uses light and dark themes, bundled Barlow fonts plus the on-device Alexa Sans, and views only: the pitch, the crests and a gradient are all built without a native graphics library.
  • Coach Studio: Vite, React and TypeScript, with an SVG tactical canvas and annotation tools.
  • Publication API: a dependency-free Node server. Workflow rules (draft, needs review, approved, published, on hold, changes requested, unpublished) live in one tested module. Staff and players get separate endpoints, and the TV can only ever fetch published reviews. The TV polls it from the simulator (10.0.2.2) and falls back to bundled demo data when offline.
  • Shared data: both apps and the server import the same typed demo content, so what the coach approves is exactly what the player sees.
  • Testing: 15 Jest tests on the TV app (access boundary, live sync, flows, privacy) and 9 workflow and API tests. We also wrote our own scripts to take simulator screenshots and send D-pad presses, so we could check every screen on the device in both themes.

Challenges we ran into

  • Getting reliable screenshots and remote input from the virtual device. There's no built-in command. On-device tools hit sandbox limits, and QEMU's screen capture came back blank. We ended up using the emulator console for screenshots and inputd-cli with KEY_* names for the D-pad.
  • Builds failing because the project lived in iCloud Drive, where the path contains a space.
  • Designing a 10-foot UI that stays simple while showing evidence, sources and voice interaction.
  • Treating responsible AI as product rules, not disclaimers: proposals stay tentative, approval is separate from publishing, and staff-only sources never reach players, including through the assistant.

Accomplishments that we're proud of

  • An end-to-end loop that really works: a coach confirms, edits, draws, approves and publishes, and the player's TV updates in seconds.
  • The player can't see anything the coach hasn't published. That rule is enforced in the data layer and the API, and covered by tests.
  • An assistant that shows where every answer comes from, and says so when it can't answer.

What we learned

  • Fire TV's strength is focused, sit-back sessions. Feedback the player studies works better there than notifications.
  • Vega's React Native layer is productive once the environment is right. Most of our time went to tooling around the simulator, which is all in the friction log.

What's next for CoachOS

  • Bee as the capture source: import the coach's match-day transcript, align it to the match clock from kickoff, link statements to moments and named players, keep them staff only by default, and file them by match. In the demo, coach statements are typed demo data.
  • Match video and tracking: moments and clips from a match camera with event tagging, and player positions from tracking data, alongside the tactical replay.
  • A real Alexa+ integration via an MCP server, keeping the same rule that it answers only from published feedback.
  • A phone companion that notifies the player when a review is published, while the TV stays the place to study it.
  • Multiple coaches and teams, roles, and authentication.

All code and assets were created during the submission window, starting 14 October 2026.

Built With

  • a-lot-happens.-coaches-scribble-notes-on-the-touchline
  • and-debriefs-fade-by-monday.-the-player-usually-gets-a-clip-and-a-one-line-note.-the-real-intent-is-scattered
  • and-it-often-seems-to-contradict-itself.-ai-can-summarise-all-of-that
  • approved-by-the-coach
  • but-a-coach-can't-send-a-player-ai-written-feedback-that-might-be-wrong.-we-wanted-feedback-that-is-specific
  • but-most-of-it-gets-lost:-notes-stay-on-paper
  • half-time-talks-are-never-written-down
  • linked-to-its-evidence
  • node.js
  • react
  • react-native-for-vega
  • say-crucial-things-at-half-time-and-clarify-them-in-the-debrief
  • typescript
  • vega-sdk
  • vega-virtual-device
  • vite
Share this project:

Updates

Submission history