Monitoring where an event is visible right away, not a day later.
We build event-detection systems on satellite imagery and UAV data: the model finds an object or an anomaly, the result goes onto a map, analytics is built around it. The value here isn't model accuracy on its own — it's how fast a human learns about the event.
Detection on imagery
Pollution Detection: oil-spill detection from satellite and UAV data. Response delay reduced from 6–48 hours to online detection.
The map as a working tool
Results on a map with history, filters and event review — so the detection can be worked with, not just notified about.
The data pipeline
Image ingestion, preprocessing, inference and result storage. Most of the engineering in monitoring lives here, not in the model itself.
Honest quality assessment
False positives and misses cost differently depending on the task. The threshold is chosen for the cost of an error, not for a good-looking metric.
Satellite or UAV data?
Both. Pollution Detection uses both sources: the satellite gives coverage, the UAV gives detail on a specific point.
Do we need labelled data?
Usually yes, at least a small set for quality evaluation. If there's no labelling at all, collecting it becomes the first step — part of the work, not a precondition.
What about false positives?
They can't be reduced to zero, so the question is framed differently: what costs more in your case — missing an event or a wasted response. The answer sets the threshold and the human review flow.
Can it integrate with our monitoring system?
Yes, that's a standard part of the job: detections are pushed to where the duty process already runs, instead of living in a separate interface.