Explore New York City through the eyes of a cyclist. Discover hidden risks, investigate near misses, and explore possible futures.
Three steps. No instructions needed.
Every ride is a 5-second GoPro segment indexed by the VAST pipeline with YOLO detections and a Cosmos caption.
Project alternative futures, expose blind spots, search similar moments, or ask GHOST for evidence.
Each tracked road user carries a live Risk Score from 0 to 100. Click it to understand why.
Everything GHOST shows traces back to a component of the VAST video stack.
78 thirty-second chunks, split into 465 five-second segments, camera nyc_bike_gopro-1.
460k boxes at 30 fps. GHOST reads them; it never adds or edits a detection.
Per-segment reasoning text and vectors power Street Memory and the Investigator.
ByteTrack-style tracking, focus-of-expansion ego motion, transparent 0–100 heuristic.
Llama 3.3 70B restates computed evidence in plain language. Deterministic fallback, always labeled.
Choose a situation. See what the AI discovers.
GHOST NYC is a bicycle safety intelligence layer built on the VAST video stack.
nyc_bike_gopro-1.Observations are YOLO boxes and tracks. Predictions are image-space projections of those tracks. Uncertainty is shown as amber. No value is a distance in meters or a measured crash probability.
Every journey reveals something different.
Each bar is one real 5-second segment, placed where it happened in the ride and colored by its peak Risk Score. Click a bar to open that moment.
Highest peak Risk Scores across all analyzed rides. Thumbnails are the actual frame at the peak.
Share of analyzed segments where each class was tracked.
Local Risk Score peaks ≥ 50, grouped by what the pipeline saw in the segment.
Segments with an estimated ego corridor (enough tracked motion) versus an approximate one.
A narrated, cinematic walkthrough of every feature on real footage, driven live inside the app. Press play, hit record.
Space pause · → next chapter · ← previous · H hide controls · Esc exit
Clean recording mode hides the controls while idle and enters fullscreen. Voices are your browser's built-in speech voices; the script below is ready for any external AI voice.
Play from any chapter. Every chapter runs the real feature on real footage, nothing is pre-rendered.
Two independent clips. Sync aligns playback position only; it does not mean the clips happened at the same time.