MomentOS is a credit-score-like system for driving, built from actual dashcam video history instead of coarse proxies.
For every trip, we use the VAST video pipeline to segment and index footage, then score risky precursors such as hard braking, cut-ins, tailgating, lane drift, pedestrians in path, and low time-to-collision. Cosmos Reason produces structured risk, cause, TTC and cue outputs. Cosmos Embed measures similarity to hazardous driving events, YOLO11 adds object proximity signals, and a W&B-hosted LLM provides an additional risk judgment.
We combine these signals into a per-clip risk score, then aggregate only driver-caused events into a long-term Safe Driving Score. Events caused by other road users are still surfaced, but do not penalize the driver.
The result is not just a number. Every score is backed by the exact video moments, detected cues, cause attribution, and a generated incident report. Drivers can review and contest events, and tier changes require human review.
On the Nexar public test split, our final blend reached
