# Overheard

- **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:** [Krutin Rathod](https://cerebralvalley.ai/u/krutin31)
- **GitHub:** https://github.com/Krut-in/overheard
- **Demo video:** https://youtu.be/aYt0TgLVfYc
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/48

Overheard turns everything a traveller hears and sees abroad (announcements, signs, departure boards, news) into the few moments that touch their bookings, in their language, with proof. Everything it ignores leaves a silence receipt with a relevance score.

Why I built it: I get anxious travelling alone about missing an announcement or taking a wrong turn, even though I'm the default tour guide for friends in NYC and elsewhere. I often see elderly people and non-native New Yorkers struggle with announcements and sudden changes. Overheard knows your itinerary and tells you only what matters.

How it works: connect Gmail (read-only) or load the demo trip. Neon pulls every booking out of the emails. Hydrogen turns them into one itinerary fingerprint: a few hundred yes/no questions in French, Dutch and English, each tagged with the booking it protects. Every input, whether an announcement, a photo of a sign or a news article, goes through one gemini-decision-1 call that answers all the questions at once and scores each booking separately, so the alert names the booking that is hit. A score of 0.8 or more alerts, 0.4 to 0.8 goes to Carbon, and anything lower leaves a receipt. The alert is translated, whispered with TTS and backed by proof: your own photo with the line circled by Nano Banana, or the quoted sentence from the source. Before the trip, a search-grounded radar sweeps French and Dutch news for strikes and closures on your dates, and rehearsal mode plays the announcements you are likely to hear. Tap "not about me" on a wrong alert and it writes sharper questions for that booking.

Results on 74 labelled items against a 6-booking Paris and Amsterdam trip: Decision 1 alone caught 15 of 18 rehearsal items at under a second median, and the full policy with Carbon on the borderline band caught 18 of 18. On 38 real news items the radar found today, it caught 3 of 3 with 5 of 35 false alerts at 0.73 s median. Keyword matching fired on 10 of 12 audio near misses. The README lists our honest failures, including a wrong-booking near tie and number-only near misses.

Built solo today: the NYC subway engine came first (0.8 s median per input), then an orchestrator session split the trip-abroad version into 8 parallel ticket sessions, with Gemini CLI writing the sample booking emails and all 36 rehearsal scripts, and Carbon reviewing every news label (38 of 38 agreed).

## More from Gemini 4 NYC Hackathon

- [INTERLOCK](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/45)
- [Glint](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/46)
- [Mozart](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/47)
- [Let There Be Sims](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/49)
- [gridlocked](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/50)
- [GridSafe (67Hz)](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/51)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/48. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
