# Uno ReVoice

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Abhishek Gaduputi](https://cerebralvalley.ai/u/Abhi7), [Ariana Daka](https://cerebralvalley.ai/u/aridaka), [Jitthen Kumarr](https://cerebralvalley.ai/u/jkumarr)
- **GitHub:** https://github.com/sairam782/Gemini4-Googledeepmind-Hack
- **Demo video:** https://youtu.be/dEFmYViQ3bE
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/35

Uno ReVoice is an AI assistant you can hand a live call to when you need a moment to think. It introduces itself, asks why the caller is reaching out, and gives you a private panel with live captions, suggested replies, and warning signs linked to what the caller actually said. You can tap a reply for ReVoice to say aloud or take the conversation back at any time.

In our demo, a caller claims to be your cousin and asks you to send $200 immediately. You tap Offer a callback, and ReVoice asks whether you can reach them on the number you already have. When the caller refuses, the panel highlights that refusal alongside the money request. You can take back the call and decide what to do next.

How we built it

Gemini Live transcribes the call while several models work in parallel. Flash-Lite drafts quick replies. Gemini 4 Hydrogen extracts facts and warning signs, Argon plans follow-up questions, and Neon reviews the conversation and checks suggestions. Each suggestion previews the exact words ReVoice would say. Once you hand over the call, you can pick a reply or let ReVoice use its highest-confidence suggestion after a short wait. Application code checks replies before speech and rejects duplicate or stale taps. After hang-up, Gemini 4 Carbon summarizes the call, and Neon drafts follow-up actions for you to approve.

Agentic self-improvement

We also prototyped an improvement loop that learns from completed calls. A Gemini 4 Neon reviewer looks for missed checks, poorly timed replies, and weak follow-up questions. It proposes lessons for a versioned playbook that guides future calls. For example, if ReVoice only offered a callback after a caller failed an identity check and asked for money, the reviewer can suggest offering a clear refusal sooner in similar calls. Application code validates proposed lessons and replays changes to risk weights against past calls. The speech and takeover rules stay under application control.

What we tested

Our scripted replay passed 62 of 64 checks across eight scenarios, split evenly between scam and legitimate calls. The checks covered risk assessment, evidence links, and suggestion coverage. The live demo shows how those pieces help someone slow down a high-pressure call and choose what happens next.

## More from Gemini 4 NYC Hackathon

- [Pyre Witness](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/32)
- [Cornell Movers](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/33)
- [SightSense](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/34)
- [GetUp](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/36)
- [Work Less Video Editor](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/37)
- [Blindspot](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/38)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/35. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
