Maritime intelligence, from AIS to the camera on the mast.

We predict from the vessel’s own history instead of its live signal. Past voyages, speed behaviour, headings, and the routes and ports that class of ship habitually uses feed one model that answers three questions at once: the most likely path from the last known position, the expected arrival time along it, and the ports it may have called at during the silence. The prediction horizon is configurable, so the same system serves a six-hour question and a six-day one.

Restoration and detection in a single pass. The pipeline removes haze, cloud and rain distortion first, then detects and classifies ships in the restored scene, with no round trip through a second system. It works at scene level rather than tile level, so a vessel straddling a tile boundary is still one vessel and its position on the chart is still right.

Range from a single uncalibrated camera, in real time, using the visual cues in the scene and what the model has learned about how each class of vessel presents at distance. The same pass detects the ship, classifies it and tracks it across frames, so one video feed produces identity and range together rather than needing two pipelines behind it.

We mine historical AIS to extract the routes that actually exist, clustered by vessel type and region rather than assumed from a chart. That becomes a live baseline. Incoming positions are scored against it continuously, flagging departures from known corridors, behaviour that does not match the vessel class, unusual clustering in fishing grounds, and route changes that are only beginning to emerge.

Anchorage detection straight from raw AIS, annotation and correction tools built for the people who do the labelling rather than the people who wrote the model, dashboards that answer spoken questions so an operator can keep their hands and eyes elsewhere, and speaker diarization over maritime voice traffic so chatter becomes searchable text attributed to whoever said it.