Skip to Main Content

PROBLEM: Investigators receive more war footage than people can sort through. The ICC prosecutor received 570,000+ files in one year; Ukraine has registered 266,000+ war crimes and convicted 317 people. Platforms strip GPS, so each clip is located by hand. One video is an anecdote: evidence needs several independent clips that agree on place and time. Fakes are mixed in: video-game clips, film scenes, and recycled footage get posted as real strikes. WHAT WITNESS DOES. Its input is video with sound. It does the sorting, not the judging: 1. Watches and listens to each clip and lists what happens, second by second. 2. Rejects fakes with four named checks, each tied to a moment in the clip. 3. Places each real clip on a map with a stated uncertainty radius. 4. Rebuilds a strike as a 3D scene on one shared clock: impact, sound wave, flags and review notes, with the source video following the playhead. Drag the timeline and everything rewinds. WHY IT NEEDS VIDEO AND AUDIO (Element 119). Light arrives before sound. The gap between flash and boom, times 343 m/s, is the distance from camera to blast. A boom that arrives together with a distant flash marks dubbed or rendered footage. Two blasts with identical waveforms are a stock sound effect. Gemini 4 Hydrogen supplies the timed events; plain code measures the audio and frames. On real-or-fake alone, transcript-only scores 38%, and five still frames match full video at 100%: sound earns its place in the timing checks. LIVE, UNSCRIPTED DEMO. A judge pastes any YouTube link or drops any video file. Six stages stream live, and a verdict (real, fake, or unsure) arrives in under a minute. Every reason is a button that jumps the video to that second. The judge then says what the clip really was, and a scoreboard keeps every answer, wrong ones included. PROOF. Scored against locations published by Bellingcat and GeoConfirmed, which are never used to locate. Dev set, n = 17: 59% of clips within 1 km (Claude Opus 5.5 on five frames: 18%; one direct Hydrogen call: 53%); median error 0.7 km (2.5 km; 0.8 km); true location inside our stated radius 76% of the time. Fakes: 100% precision and recall, a tie with every baseline. For two rebuilt Kharkiv strikes, our impact point is 114 m and 70 m from Bellingcat's. Dev and test clips were split by hash before tuning; the test set runs once. WHERE IT BREAKS. A YouTube link is scored from a written description, so its score varies between runs and gets a wide "unsure" band. Most footage is aftermath, so a flash-to-boom distance is often unavailable. The 3D city is built from OpenStreetMap footprints, not a scan. The two incidents on screen were assembled by a person; automatic grouping is scored (82%) but not displayed. SAFETY. No blame, no legal verdicts, no weapon named as fact. Clips showing military objects are never located. All product code was written during the event; nothing was trained. It sorts. Humans decide.

Team

kevinyyinjamwan

More from Gemini 4 NYC Hackathon

We design an agentic system to automatically create + manage intelligent AI opponents/NPCs in video games. The suite of agents analyzes user gameplay…

Adaptive NPCs project preview

Manually routing PCBs is time-consuming and cost-inefficient. Current best algorithms (freerouting, PCBRouter) score very low on PCBWorld's hardest PCB…

routr project preview

Heliocam is a camera app where Gemini controls the camera. It watches what you're pointing at, sets the ISO, shutter, focus, white balance and zoom for you…

heliocam project preview

Repairs and assembly fail because manuals are lost, vague, missing, or nonexistent, so people break things or throw them out. KLIK turns a photo of any object…

Klik project preview

Iron Pulse turns your laptop into a 3D boxing ring. Fight an adaptive AI rival in real time using just your webcam. Throw real punches and lean to dodge…

Iron Pulse Arena project preview

Text your raw footage to an iMessage number and get a finished, trend-informed Reel back in the same thread.

Sticky project preview