# Soul

- **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:** [Harshit Aggarwal](https://cerebralvalley.ai/u/harshitaggarwal)
- **GitHub:** https://github.com/harshitaggarwal01/splatdiff
- **Demo video:** https://youtu.be/wfl40iCmrsM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5188

Second Look tells a family what changed in an aging parent's room between visits — with proof.

The problem: about 1 in 4 adults over 65 falls each year (CDC), and home hazards are one factor families can actually fix. But the person who visits isn't always the person who worries, and nobody remembers exactly where the walker stood three weeks ago. A walker pushed out of reach or a bag left in the path goes unnoticed.

How it works: film a 20-second walkthrough of the room when it's set up safe (e.g. after a hospital discharge), and again on a later visit — anyone can film. Second Look blurs people and papers, inventories the room, tracks every object through both videos, and reports what was moved, added, removed or left on the floor. High-risk items come first, each with plain "Do this next" steps (ask Dad first; never ask him to move furniture alone).

Why it's trustworthy: every finding carries its evidence — SAM 3.1 boxes, frames and confidences from your own videos — and anything the reasoning model claims that the evidence doesn't support is struck from the report, in front of you, with the reason. On real phone footage of a computer lab, it caught chairs pulled into the aisle and struck a "trash bin added" claim because the first video never showed that corner.

Built for families: a care watchlist (typed or spoken); Dad's own words ("I moved the lamp to read") mark a change as his choice, not an alarm; a live words-only family feed with an agent that answers "Where was his walker last seen?"; updates written for someone nearby, far away, or Dad himself; per-room baselines and reminders; and a glasses simulation where Dad wears the camera and it remembers where he left things. Private by construction: no installed camera, blur before reasoning, raw videos deleted by default.

Honest numbers: on our one scoreable test room the raw model invented 25–40% of claims; grounding struck every one — a smoke test, not a benchmark.

Live demo: https://harshitaggarwal01.github.io/splatdiff/

## More from Meta Global AI Developer Hackathon

- [Centaur Video](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5185)
- [Honeymeta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5186)
- [Teamllamainspector](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5187)
- [BakedBot AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5189)
- [Cicada](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5190)
- [The CareThread](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5191)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5188. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt