Attach a model to a feed and describe the alert

Point the platform at any camera, drone feed, or video stream. The model attaches to the feed and inference starts the moment the stream connects, every camera watched in parallel.

Write the rule the way you would brief a person: 'Alert me when rust or corrosion appears on the line.' The condition compiles against your model's classes - no query language. Per-class confidence thresholds ship as working defaults and stay yours to tune; the field guide's confidence lesson covers how to set them against your real base rates.

Slack, email, or your own API endpoint. Every alert carries the evidence frame with the detection drawn on it, so the person paged can judge in one glance, from a phone, without logging in.

tip · Route by severity: routine counts to a channel, critical detections to the person on call.

Detections rank themselves by the model's own confidence. Under the thresholds you set per class, routine events file automatically, borderline cases queue for human review, and critical classes page immediately. The review queue is also a labeling surface: every triage decision is a verified label.

Every triage decision is logged against the model's prediction, and the override rate is the drift signal the loop acts on: when corrections cross the retrain threshold you set, a retrain is scheduled automatically, built from the flagged evidence frames your reviewers already judged - exactly the ones the model got wrong. One habit before the threshold trips: if overrides surge suddenly, check the drift catalog first - an abrupt cliff usually means a pipeline event, and a revert beats a retrain.