Everything here is a pilot acceptance criterion — measured by your engineers, on your equipment, at your resolution — not a datasheet specification.
At 300 km/h the aircraft covers 83.3 m every second. At 30 processed frames per second that is 2.8 m between looks; at 5 fps it is 16.7 m. The gap between frames is distance the track has to bridge with no new information — no model compensates for it.
300 km/h = 83.3 m/s → ÷30 fps = 2.8 m · ÷5 fps = 16.7 m.
Range is geometry: it is set by how many pixels land on the target, not by TOPS. On a 4 m target from the full frame the model has enough pixels out to 2,750 m. A weak board downscales the frame to a 640 input to keep up — and the same target is gone by 917 m.
From the target entering the field of view to a confirmed event on board: 247 ms when the chain closes on the aircraft, 794 ms at 7 fps. Routing the same decision through the ground adds about 200 ms on the radio link alone — a link the aircraft does not need.
Frame rates are indicative single-stream figures for YOLO-class models. Latency and autonomy are measured on your equipment, not declared.
hardware decode of the full-resolution frame
your model, on every frame
held through the manoeuvre and after loss
across the sensors on the platform
confirmation on three frames of five, not one trigger
written on board, synced when the link returns
nothing leaves without your say
Every aircraft and every model version, tracked.
Staged rollout to a subset first, then the fleet, with rollback.
Detections and video kept for retraining.
The next version trained on your own archive.
Not writing a detector, but keeping one alive across sensors, seasons, theatres and board revisions. We run all ten with your team, on your data, and the output at every stage belongs to you.
Curated clips around every detection plus a hard-negative sample, recorded on board and synced when the link returns.
Labelling against your classes and your confidence policy, to an agreed quality bar.
On your own data, fine-tuning from your checkpoints, runs reproducible by your engineers.
Held-out evaluation on your scenes, so the accuracy you publish is a number you measured.
Quantisation and compilation for the runtime on the module you chose, without rewriting the model.
Measured on the target compute, not a datasheet — frames, latency and occupancy at your resolution.
Signed model bundles with the configuration that produced them, versioned and reversible.
Over the air to a staged subset first, then the fleet. No re-flash, no recall, no aircraft recovered to update it.
Drift and accuracy watched per model version, every detection tagged to the build that produced it.
The archive feeds the next version. The loop closes without anything leaving your infrastructure.
Events to the archive, archive to retraining, retraining back to every aircraft over the air — and the weights at the end of it are yours to take anywhere, including away from us.
We qualify the board you pick and compile your model for its runtime — we do not resell computers. Hardware sits outside the licence: about $1,109–1,214 per aircraft at July 2026 prices, bought by you.