Traffic cameras you already own, turned into evidence you can act on.

TrafficAI watches live public traffic cameras, detects risky behaviour, ranks every camera by measured risk, and writes dispatch reports that must pass a hallucination gate before anyone reads them.

View the live risk ranking

Currently ranked highest: RT 50 AT PATRICK HENRY DR. at ELEVATED (risk index 2.9). One of the 6 was refused. Read the reports.

The record so far

Cameras discoverable

1,677

unique camera ids in the current Virginia DOT feed snapshot, sanitized at discovery

Soak runs completed

3

multi-hour captures on live cameras across afternoon, rush hour and overnight conditions

Alerts raised

8,360

near miss, tailgating, wrong way and pedestrian intrusion

Human verdicts

815

labelled evidence snapshots that set every precision number

Scene descriptions

1,268

written by a vision language model off the frame loop

Automated checks

108

four verifier suites, all passing when this page was built

Each figure names its source. The full measurement story is under measurement.

What the system does

1. DiscoverThe public VDOT feed yields 1,677 live camera streams. Every URL is sanitized and checked before it is opened.
2. PerceiveDetection and tracking run on every frame. Each camera learns its own traffic flow and roadway.
3. AlertNear miss, tailgating, wrong way and pedestrian intrusion checks fire only past gates built from human labelled evidence.
4. DescribeA vision language model narrates alert scenes without slowing the frame loop.
5. RankRisk per 100 vehicles observed, weighted by measured precision and the road conditions of that hour.
6. ReportNarratives are assembled from retrieved events and validated. A report that fails is quarantined, never dispatched.