# Sightloop

- **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:** [Deekshith H U](https://cerebralvalley.ai/u/deekshith1304)
- **GitHub:** https://github.com/Deekshith9238/sightloop
- **Demo video:** https://youtu.be/FGRjTV1BF_4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2928

SightLoop shows what covers a web page on a phone, the way a visitor arriving from Google sees it, and proves a fix works before you ship it.

The problem: a phone visitor taps a search result and gets a full-screen "Download our app" sheet, a lead form or a chat pop-out over the page. Many just go back. Google asks sites not to cover content with intrusive interstitials, but gives no number and no way to see your own page as that visitor does. Small Indian businesses (AC repair, home services, tour operators) mostly find out from lost calls. In a scan of 165 pages from real Indian service searches, 13% (95% CI 8–18%) had an overlay covering 20% or more of the screen.

How it works: paste any link. SightLoop opens it on the 8 phone screens Indians use most (the top six resolutions from StatCounter, plus a large iPhone and a foldable that opens mid-visit), with a Google referer, and records the first 12 seconds. The page code proposes a box for every overlay; Meta's SAM 3.1 tracks each one frame by frame, forwards and backwards in time, giving exact coverage per frame and the moment it appeared. Muse Spark Open labels each overlay (app install, newsletter, chat widget, consent banner, site header). The verdict is computed in code: a non-consent overlay covering 20% or more of the screen is flagged. Muse Spark explains the finding in plain words and writes a CSS fix.

Then the loop: SightLoop injects the fix, records all 8 phones again, and checks. A fix only counts if the page still works: its own button must be usable afterwards, so "hide everything" fails. If it fails, Muse Spark gets the failed numbers, the last frame and the HTML of what is still on screen, and writes a new fix (up to 3 attempts). If a fix passes only by removing the popup, it asks once more for a small banner that keeps the popup's button.

Example: on NoBroker's AC repair page an app-install sheet covers 50% of the screen 2.5 s after arrival and dims the page, hiding the BUY button. The first fix removed it; the second kept it as a small banner with its button: 10%, verified on 8 of 8 phones.

Evaluated honestly: I labelled 30 sites from the recorder's frames before seeing any tool output. On the test set, the verdict agrees with my labels on 91% (95% CI 77–100%); the repo has the per-site results and a failure gallery.

## More from Meta Global AI Developer Hackathon

- [Hawkaii](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2925)
- [DeckTwin.com](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2926)
- [Hallucinated Dev](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2927)
- [Abu Bakar Siddik](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2929)
- [Cortraders](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2930)
- [ProofLoop_enhanced](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2931)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2928. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt