# Screen-Time

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Aldo tile](https://cerebralvalley.ai/u/aldo311)
- **GitHub:** https://github.com/aldot311/screen-time
- **Demo video:** https://youtu.be/WTzZLN3Yb_M
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2441

A brand pays a creator, say $300 for a Reel, on terms like "show the can for 3 seconds, say the brand name, disclose the ad". Today someone watches every Reel to check that, or nobody does. Screen-Time does the check and shows the moment that proves each term.

The brand writes its terms in plain English and sends each creator a link. When a Reel comes in:
- SAM 3.1 (Meta Model API) tracks the product through every frame of a 15 fps copy, from a text prompt.
- Muse Spark compares crops of every tracked object with the brand's product photo and rejects look-alikes, such as a different LaCroix flavor or another brand's can.
- Muse Voice Transcribe, given the brand names as keywords, finds where the brand is said and whether the creator says it's an ad. Muse Spark also reads burned-in "#ad" text.
- Muse Spark turns the plain-English terms into typed rules, checked in code: seconds on screen, first seconds, one unbroken shot, share of frame, brand said, ad disclosed.
- The report gives pay or hold for the creator's fee, with a play button on each term that jumps to the proof. Close calls are flagged "worth a look".
- A follow-up agent (Muse Spark with 6 tools) drafts the note to a creator who fell short and queues a re-cut request. It cannot move money: pay is refused in code unless every term passed, and sending a link needs the brand's approval.
- The campaign page totals pay and hold across creators, with a CSV export. Every Meta call is itemized on a receipt ($0.11-0.23 per Reel, an upper bound).

Accuracy, measured on 11 Reels as delivered: per-frame precision 98.4%, recall 91.1%, against labels drafted by two AI vision passes and checked frame by frame by a person (11 of 11). Method, confidence intervals and the cases it gets wrong: https://screen-time-inky.vercel.app/accuracy

Live product (no signup): https://screen-time-inky.vercel.app (start at /judge, "How to judge this in 5 minutes")

Built during the hackathon: all of it, first commit Oct 28, 2026. Built by Aldo Tile with Claude Code as the coding assistant. The LaCroix campaign is a test run with friends, who agreed to be shown; not affiliated with LaCroix or Poppi. The demo video's voice is AI (ElevenLabs via fal.ai) and the music is Google Lyria.

## More from Meta Global AI Developer Hackathon

- [Avocaders](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2438)
- [TBone](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2439)
- [Prince Tiwari](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2440)
- [S47](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2442)
- [Latent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2443)
- [Snaprint](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2444)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2441. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt