Solutions · AI Use Cases
On-device vision jobs — driver monitoring, drive assistance, worksite safety, asset security — fused with live vehicle telemetry. Camera and NPU share memory, no CPU pixel copy. Bring an ONNX or TFLite model; up to ~100 TOPS on the AI-class tier.
Catalog of AI use cases
Each job is a model on the device fused with synchronised OBD, GNSS, and CAN. Video is analysed on the box — only events and the clips you choose ever leave the vehicle. Bring your own model or have one trained from your footage.
Jobs run on detection and classification today. Jobs marked on study use pose, distance estimation, or text reading and are in development — validated per fleet before they ship.
Driver — facing in
Cabin-facing vision watches the driver, not the road. Every detection is gated by trip, speed, or ignition state, so alerts fire in context — and the in-cab feed is analysed on the box, never uploaded.
Eye-closure and head-nod patterns are scored on the in-cab optic and weighted by how long the driver has been on the road, so a long night shift raises sensitivity. Inference runs on the box — the cabin feed never leaves the vehicle.
Head-down, eyes-off-road posture is classified continuously, but the alert is held until GNSS speed crosses a moving threshold — a parked glance at the phone does not fire it. On-device only.
Hand-to-ear and handheld-device gestures are classified while the vehicle is in motion, separating a quick glance from sustained handheld use. No frames are uploaded to make the call.
Seatbelt-on state and in-cab policy breaches are read from the image and correlated with ignition and motion, so the check applies only once the vehicle is actually in service.
Road — facing out
Road-facing vision watches the path ahead. Detections fuse with own speed, steering, and turn-signal CAN, so a warning reflects what the vehicle is actually doing — not a raw pixel guess.
The lead vehicle and its closing gap are tracked on the forward optic and fused with own speed, so the imminent-impact warning scales with how fast you are actually approaching.
Lane drift is detected on the road optic and suppressed the moment turn-signal CAN state shows the manoeuvre is intentional — signalled lane changes do not nuisance-trip.
Pedestrians and cyclists in the forward path are detected and ranked by speed and steering angle, so the closest in-path risk is surfaced first.
Following distance is estimated from the forward image and graded against speed for a time-gap score. Distance estimation is in development and validated per fleet before it ships.
Speed-limit signs are read by on-device OCR and compared with measured speed to raise an over-speed event. Sign reading is in development and validated per fleet before it ships.
Worksite — around & behind
Around-and-behind vision covers the low-speed danger zone — reversing, loading, and manoeuvring near people. The safety alarm fires locally on the box, with no cloud in the loop.
A person entering a defined danger zone around a machine or reversing vehicle is detected and fires a local alarm at the box — no cloud round-trip in the safety loop.
People and obstacles alongside or behind are detected during low-speed manoeuvres, covering the blind spots a mirror misses.
Pose estimation flags a person at the edge of an open liftgate during loading. Pose work is in development and validated per fleet before it ships.
Before the vehicle is cleared to move off, the rear area is confirmed clear of people. Gear-gated and in development, validated per fleet before it ships.
Asset — cabin & cargo
Cabin-and-cargo vision watches the asset when the vehicle is parked or moving unattended. Detections are stamped with GNSS and gated by lock, ignition, and motion state.
A passenger, child, or object left in the cabin is detected after ignition-off and door events, so the alert fires once the vehicle is parked up and emptying.
A person entering the cargo area while the vehicle is parked and locked is detected and stamped with GNSS location for the event record.
An open door read from the image is cross-checked against motion telemetry, so the event fires only when a door is open while the vehicle is actually moving.
On-demand & bring-your-own model
The runtime is model-agnostic. Ship an ONNX or TensorFlow model and it is converted to the on-device format for you, or have one trained from your labelled footage. Models load when a pipeline starts and release when it stops, so one device can carry several use cases from the catalog above.
ONNX · TFLite Bring an ONNX or TensorFlow model; it is converted to the on-device .tflite format for you. Up to ~50 MB; under 5 MB keeps latency lowest.
on-device NPU Frames are analysed on the NPU before a byte leaves the box. Raw video stays local — only events and the clips you choose to upload ever leave the vehicle.
load · unload Models load when a pipeline starts and release when it stops. Run a single model or cascade several across the NPU on the same frames.
detection + telemetry Every detection is fused with synchronised OBD, GNSS, and CAN — so a use case can be gated by speed, gear, or door state. The full pipeline lives at AI Funnel.
Where it runs
FAQ
Yes. Ship an ONNX or TensorFlow model and it is converted to the on-device .tflite format for you, up to ~50 MB. Under 5 MB keeps latency lowest. Or have one trained from your labelled footage.
On the device NPU, on AI-class silicon. Raw video is analysed on the box — only events and the clips you choose to upload ever leave the vehicle. There is no CPU inference fallback.
Yes. Models load when a pipeline starts and release when it stops, so one device can carry several jobs from the catalog — run a single model or cascade several across the NPU on the same frames.
Cloud retraining, quantisation, and OTA model delivery are managed by the AI Funnel pipeline. Retrained models land on the device without a full firmware update.
Pick the jobs you need, bring your own model or footage, and we will sketch the on-device pipeline during the call.
One reply from engineering, ~24h. No deck, no NDA.