Predictive Fleet Analytics Platform for Mer**** Logistics, Midlands UKCase Study

The operator had years of telematics and job history sitting unused. Customers were increasingly asking for accurate delivery windows, and the traffic office was giving estimates based on experience, which meant confident answers that were often wrong. Breakdowns were the more expensive problem: vehicles were maintained on fixed service intervals regardless of how hard each unit had actually worked, so some were serviced unnecessarily while others failed on the road, where recovery and a missed delivery cost far more than a workshop slot. Seasonal capacity planning was similarly instinctive — the business knew certain periods were busier but could not quantify by how much or which lanes were affected. The data problem underneath all of this was that telematics arrived as raw position pings with no job context, so a vehicle's movement could not be tied to the delivery it was performing. Without that link, none of the predictive work was possible. There was also justified scepticism internally, since the planners had seen confident software before and did not trust a model over their own judgement.

The first phase was not modelling but data engineering: we normalised the telematics feed into a job-aware event stream so every position ping could be attributed to the consignment it belonged to, which is what made everything downstream possible. Predictive work then runs as a Python service. Delivery ETAs are predicted from historical journey times on comparable lanes, time of day and live traffic from Google Maps Platform, rather than from a static distance calculation — the model learns that a given lane behaves differently at different times. Maintenance prediction uses engine hours, load profile and fault codes from the telematics feed to flag vehicles whose usage suggests attention before the fixed interval falls due. Demand forecasting projects capacity requirements by lane and period so the operator can plan subcontractor cover ahead rather than scrambling. Results surface in a React dashboard where every prediction shows its supporting history and a confidence indicator, because the planners were never going to trust an unexplained number. The whole platform runs on AWS with the model service isolated from the web application so a long training run never affects the operational dashboard. We deliberately kept predictions advisory rather than automated. Related: AI & ML development and our logistics industry practice.