AI Camera Intelligence That Turns Footage Into Reviewable Events

The in-cab feed is the sensitive one and it is treated as such. Camera Intelligence is designed to surface defined safety events — phone use, eyes off the road — as short clips, not to stream a continuous view of a person at work to a supervisor's screen. Which cameras are active, whether the cab-facing feed is enabled at all, and who can open a clip are configurable per fleet, so the deployment can meet whatever consent process and workplace policy you operate under. We would always advise telling drivers plainly what is captured and why; a camera programme that feels like surveillance gets sabotaged, taped over or fought, and stops producing usable evidence.

Not necessarily. The module is built to ingest feeds from cameras that already expose a usable stream, and many AI-dashcam and MDVR devices in the market do. What matters is that the device can deliver frames at a workable rate and be paired to the vehicle's tracker so clips inherit location and driver. Very closed devices that only record to a local card and never release the feed are the genuine blocker — those either need replacing or can only be used for after-the-fact retrieval, not live detection. We will tell you honestly which of your existing hardware qualifies before anyone buys anything.

It will, and the whole workflow is built around that fact. Sun glare, a passenger's phone, a legitimate emergency stop and heavy market traffic all produce clips that a model might flag and a human will dismiss in three seconds. That is exactly why nothing here auto-penalises a driver: a computer-vision flag is a candidate for review, not a confirmed event. Every dismissal also tunes the system — triggers that repeatedly fire on nothing get their thresholds tightened for that fleet — so the review queue gets shorter and more trustworthy over time rather than drowning supervisors in noise.

Those two work from motion and telematics data — accelerometer, GPS, speed — to score how a vehicle is being driven and rank drivers over time. They are excellent at the 'how much and how often' question and need no camera. Camera Intelligence adds the layer they cannot reach: the visual evidence and the events only a camera can see, like a following distance that collapses or a phone raised to the ear before any harsh input registers on a sensor. In practice they work together — a hard-braking spike on the scorecard becomes far more useful when the clip beside it shows the auto-rickshaw that caused it. This module supplies the video; the scorecard turns confirmed events into a trend.

By design, not much of it. The point of the module is to avoid warehousing continuous video — computer vision runs over the feed and keeps the short clips around detected events, so what is retained is a handful of bounded incidents per trip rather than the whole shift. Retention windows for those clips are configurable to match your policy and any regulatory requirement, and clips tied to an open claim or investigation can be held beyond the default while the rest age out. Continuous recording, where a fleet specifically wants it, is a heavier storage decision we scope separately rather than switch on by default.

A human reviews. The model's job ends at flagging a candidate and cutting the clip; the decision to confirm it, dismiss it, or escalate it to coaching or a claim is made by a person looking at the footage. Nothing about a driver's record changes on an unreviewed AI flag alone. This is deliberate — both because computer vision gets things wrong often enough that acting on it blindly would be unfair, and because an evidence trail that stands up with HR or an insurer needs a named person's judgement attached, not just a model's output.