One place to plan, stream and learn from every live event

Hackathon build: new work since Oct 26, 2026

Judges start here. Everything below was built in the hackathon window: commits dated Oct 27, Oct 31, Nov 1, or Nov 2, 2026 — nothing backdated. Verify it yourself: git log --since=2026-10-26 reproduces the full in-window history (67 commits at time of writing). Nothing else in this demo is claimed as new.

Built with Muse Code (the agentic coding CLI, powered by Muse Spark): the agent wrote the code, ran the checks, and iterated against failing tests. Models via the Meta Model API; hosting on Cloudflare Pages + Functions with Workers KV, D1, and R2. Built on a 2015 13-inch MacBook Pro.

History mechanics, not features: 41786a6 reverts main to 034f158; ff34f64 then re-applies the Day 2–7 work. Not claimed: the Oct 15–19 base (rebrand, login shell, nav rail, seeded demo content) and any view untouched by the commits above. The same map lives in the README build map.

Who built it

What you’ll see

Every screen runs on the same live data, and Muse can answer questions about any of it.

The job, and where it lives in the app

What a streaming and broadcast engineer owns on a live event, mapped to the part of Muse Live that supports it.

Architecture

A static single-page app at the edge, a thin server layer for anything secret, and an agent loop where the model plans and the browser runs the tools.

On your face: the Meta Ray-Ban Display companion

The producer's view on Meta glasses — tally, alerts, and show status, live from the same backend as the dashboard.

1. The desktop publishes a live status every 10 seconds: every show's state and health, open incidents, and what's next on air — or "unknown", never an invented next item.
2. The glasses app (/glasses/, built on @wearables-ui-toolkit/mrbd) renders that status with the same sign-in as the main app.
3. Publishing runs on a background cadence, so switching browser tabs never silently turns the glasses into a demo. If the status goes stale for over 60 seconds, the glasses label it "Demo data" instead of passing old numbers off as live.

How Muse answers a question

The model never sees the database and never writes on its own.

1. The browser builds a live snapshot: every show’s readiness, dates, stream health and open incidents.
2. POST /api/museai adds the persona and tool schemas and makes one call to Meta Model API, so the key never leaves the server.
3. If the model asks for tools (production, techbook, crew, schedule, streaming, incidents, weather…), the browser runs them against the plan of record and posts the results back. This repeats for up to 8 steps.
4. Anything that would change data comes back as a Confirm card. A person decides, and Undo is offered.

How Meta AI was used

Three models, three distinct jobs — and a deterministic fallback that is never a silent substitute.

• Muse Spark 1.3 — the Muse drawer: reasoning and tool calls over the live production snapshot (POST /api/museai). Every write returns as a Confirm card; the model never writes on its own.
• SAM 3.1 — Stream Check frame segmentation (lower-third caption bar concept). Black frames are caught by pixel math, and any check that errors reports unavailable, never clean.
• Muse Voice Transcribe 1.0 — chunked live captions, started manually only (playback never auto-starts billing), with a searchable transcript and SRT/VTT export. Failure shows “captions unavailable”, never invented lines.
• Show Caller itself is deterministic rules, not a model call: it reads schedule, streams, incidents, crew, and equipment into a ranked brief with the source record behind every line, proposes a fix, and executes only after human approval re-verifies the world.
• No image-generation model is called anywhere in this codebase; AI-generated survey photos are labelled illustrative in the app. If the Meta Model API is unreachable, built-in rules answer instead — degraded, never silent.

How it’s checked

From the test and eval suites in the repository and an axe accessibility audit.

Decisions and trade-offs

Honest limits

• Viewer counts, bitrates and other pipeline metrics are simulated unless a show is connected to real telemetry. Every simulated number is labeled “Demo”.
• Stream monitors play public HLS sample streams, standing in for live program feeds.
• Most views are string-template JavaScript from an earlier build. Newer modules are self-contained, and docs/MIGRATION.md plans the move to React and TypeScript.
• Shared storage is last-write-wins per key. That is fine for a small team; a real deployment needs per-record merges or a database.
• This is a concept demo, not an official Meta product.

Privacy

• Google Analytics (G-35PR2RKJ51) is consent-gated: analytics storage defaults to denied and the tag is skipped entirely on localhost / 127.0.0.1. No analytics hit is sent until consent is granted.
• First-party product telemetry (js/muse-observe.js → /api/observe) runs only for signed-in sessions. It counts errors, views and Muse questions (never question text) into daily aggregates (KV, 120-day expiry); admins read them at /ops/health.html.
• Audio for captions and frames for Stream Check are sent to the Meta Model API only when you use those features. See Third-party notices.
• The demo video above loads nothing from YouTube until you press play: the thumbnail is a static image, and the player (via youtube-nocookie.com) is created only on click. Playing it is subject to Google's Privacy Policy.