The customer operated a regional logistics fleet of approximately 240 vehicles across multiple ASEAN countries, with a sustained safety challenge: a year-over-year incident trend that was effectively flat despite repeated investments in driver training, route optimization software, and conventional dashcam programs. Their existing visibility stack was, on paper, comprehensive. In practice, it was generating data that no one was acting on.
Over a structured eighteen-month engagement, we deployed an integrated ADAS-plus-video architecture across the fleet, paired with a Singapore-coordinated operations layer that turned per-vehicle telemetry into reviewable safety events. The deployment produced material reductions in both incident rate and severity, and — perhaps more importantly — a sustained behavioral shift in driver risk profile that has persisted through the post-engagement period.
The architecture.
Three components, deployed in sequence:
Vehicle layer — ADAS-integrated multi-camera.
Each vehicle received a forward-facing ADAS unit (lane departure, forward collision warning, following distance), a driver-facing camera with on-device fatigue and distraction inference, and a rear/cargo-area camera. Importantly, the cameras were not isolated dashcams — they fed a single edge inference module that correlated events across all three feeds in real time.
Edge inference — on-vehicle correlation.
The correlation matters. A forward collision warning paired with a detected driver-distraction event in the preceding three seconds tells a different operational story than either signal alone. The on-vehicle module did this correlation locally — generating high-quality, low-volume safety events rather than the raw stream of dashcam triggers that typically overwhelm fleet safety teams.
Command layer — coordinated review.
Events flowed into our Singapore command layer, which combined them with route data, driver records, and customer-side operational context. The customer's safety team received a daily prioritized review queue rather than a continuous firehose of triggers. Coaching conversations became structured rather than reactive.
Twelve-month outcomes.
The incident reduction held across both at-fault and not-at-fault categories, which is the indicator that genuinely matters. Many fleet ADAS programs reduce at-fault incidents while leaving not-at-fault flat or rising. Reductions across both categories suggest a meaningful shift in driver situational awareness, not just in their compliance with measured behaviors.
The coaching acceptance metric — the percentage of driver-feedback sessions that the driver rated as fair and actionable post-session — was the lead indicator we watched most closely. When this number is rising, the safety culture is consolidating. When it stagnates or falls, the program is generating data without changing behavior. A sixty-two percent increase over twelve months is in our experience a top-decile result.
The single best predictor of fleet safety outcomes is not the technology that detects events. It is the quality of the conversation that follows the events. // SG-COMMERCIAL · FLEET PRACTICE
What worked, and why.
Three architectural decisions accounted for most of the result:
On-vehicle correlation, not cloud correlation. Earlier iterations of similar programs streamed raw camera data to a regional cloud for correlation. That model produced higher false-positive rates, higher bandwidth costs, and a meaningful lag between event and notification. The on-vehicle correlation model is faster, cheaper, and produces cleaner events.
Structured review, not continuous monitoring. The customer's previous program had safety analysts watching live feeds. This is operationally seductive but practically ineffective — humans do not maintain attention across 240 video streams. The structured-review model presented analysts with a prioritized queue of correlated events, with the supporting video pre-fetched for context. Throughput per analyst rose by roughly four times.
Driver feedback within twenty-four hours. The closer the coaching conversation sits to the underlying event, the more credible the conversation is to the driver. Programs that surface events five days after the fact produce defensive drivers; programs that surface them next-shift produce engaged ones. Our deployment averaged eighteen hours from event to coaching conversation.
What did not work — and what we changed.
The first three months produced disappointing results. Driver pushback was high, false-positive rates were higher than projected, and the safety team was overwhelmed by event volume.
Three changes corrected the trajectory. First, we re-tuned the inference thresholds against the actual driving conditions in the customer's territories — the out-of-box thresholds, calibrated against European driving patterns, generated too many events on ASEAN urban routes. Second, we introduced a per-driver baselining model that suppressed alerts for behavior consistent with that driver's safe-baseline pattern. Third, we restructured the coaching workflow to be peer-led rather than supervisor-led, which dropped defensive responses significantly.
These changes took roughly four weeks to implement and another four weeks to validate. The lesson — for ourselves as much as for future customers — is that the first ninety days of a fleet AI deployment are a tuning phase, not a steady-state phase. Customers and integrators who commit to the tuning produce results. Customers who expect steady-state from day one do not.
Closing.
ADAS-integrated video is now a meaningful product category for Ubitron Global, with active deployments across logistics, public transit, and managed-fleet operations. The architectural pattern described here — on-vehicle correlation, structured review, fast feedback — is repeatable across vehicle classes and operating regions. The economics work because the safety outcomes are durable. We are happy to share the deployment model in more detail with operators evaluating similar programs.