# SwayamVGoswami

- **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:** [Swayam Goswami](https://cerebralvalley.ai/u/Swayyyyy_02)
- **GitHub:** https://github.com/Swayam13-exe/VoltProof
- **Demo video:** https://www.youtube.com/watch?v=yxoXfdpyZfo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4101

VoltProof AI is an AI operator for solar-plus-battery sites such as clinics, cold stores and homes. You type a goal in plain language (tested in English, Spanish, Hindi, Portuguese and German), for example: "Planned outage from 9 AM to noon. Keep the vaccine fridge on, and from noon until 8 PM keep at least 40% in the battery."

Muse Spark Open turns the goal into a short, checkable list of rules. A linear-programming optimizer plans the 24-hour battery schedule. A physics verifier, two more deterministic checks and an independent AI reviewer must pass before the plan is accepted. The plan is then replayed hour by hour on the same-day weather and compared with a rule-based battery. The AI proposes; deterministic tools decide.

Why: a plan written as text can quietly break physics, for example charging past full or counting on energy that is not there when the grid goes down. So the AI never writes setpoints. A failed check goes back to the agent with exact reasons, for at most two repair rounds; a plan that still fails physics or coverage is escalated to an operator instead of guessed.

Results (VoltProof Bench, 260 episodes, 3 sites, 5 languages, one site held out, everything frozen before the test run): goals met 98.8% against 90.4% when the AI writes the schedule itself (+8.5 points, 95% interval +5.0 to +11.9). On outage days 45 of 45 VoltProof plans passed the physics check, against 8 of 45 for AI-written schedules. On 101 harder goals: 94.1% against 88.1%.

The demo also reads tariffs from a photo of a bill (the user confirms every number) and has approve and reject buttons.

Limits, stated plainly: weather is a same-day model analysis, not a measurement; sites, tariffs and bills are synthetic; no real device is controlled; translations were not reviewed by native speakers; real bills are untested.

## More from Meta Global AI Developer Hackathon

- [yamagh](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4098)
- [Crystal Palace](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4099)
- [AMMA LABS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4100)
- [Yonro](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4102)
- [Solo Anchor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4103)
- [Kintsugi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4104)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4101. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt