# Meta IntentLoop

- **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:** [Akshay Kumar](https://cerebralvalley.ai/u/iakshay)
- **GitHub:** https://github.com/Akshay1018/meta_ai_hackathon_intentLoop
- **Demo video:** https://www.youtube.com/watch?v=UAw3k_FNUm8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/591

IntentLoop is a session layer built during a hackathon on a faithful mock Instagram UI. It is a hackathon prototype: not the production Instagram app and not connected to Instagram’s API.

We all know that Instagram already helps people manage time (Time spent, daily limits). When a limit is reached, users know how long they were on the app—but not what this visit was for, or whether they completed the reason they opened it (for example, “one DM” during a call, then Reels). That gap often produces guilt rather than closure.

IntentLoop preserves familiar patterns—feed first, no intent quiz on launch, the same daily limit sheet—and adds one action at the limit moment: See visit summary. One tap produces a structured visit summary: inferred intent versus drift, time + where (minutes per surface), a plain outcome line, and optional next steps. Summaries and visit history remain on the user’s device in this prototype.

Meta Model API (live on Vercel)
We use specialized models for distinct jobs—not a single chat thread:
• Muse Spark 1.3 — structured JSON for IntentCard (DMs/Create), ExitBrief (visit summary), vision, voice follow-up, and surroundings copy
• Muse Voice Transcribe — short voice recap on the summary sheet
• SAM 3.1 with Spark (POST /api/segment) — opt-in Reels record surroundings: with consent, the camera preview is analyzed; segmentation and Spark produce one calm on-screen nudge (for example, another device in frame)
• Muse Image — optional private visit card

Intent is cached per surface per visit; models run at meaningful moments (intent surfaces, limit → summary, and opt-in extras), not on every scroll. Time + where refresh on Your activity updates local tab timing; it does not invoke the Model API on each refresh.

You can check the links below.

Live demo: https://meta-ai-hackathon-intent-loop.vercel.app/

Primary submission demo (~3 min, narrated): https://www.youtube.com/watch?v=UAw3k_FNUm8
Silent UI walkthrough with all features (~2.5 min, screen only): https://www.youtube.com/watch?v=hzwBpWJo5Xw

## More from Meta Global AI Developer Hackathon

- [StationSync LLC](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/588)
- [Suite Fleet](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/589)
- [Blackout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/590)
- [athlete2humanoid](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/592)
- [ZeroLeft](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/593)
- [Veil Privacy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/594)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/591. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt