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

Customers were asking for accurate delivery windows, and the traffic office was answering from experience. Those answers sounded confident but were often wrong, which eroded trust with shippers. Maintenance was the costlier issue: vehicles were serviced on fixed intervals regardless of how hard each unit had worked, so some went into the workshop unnecessarily while others failed on the road, where recovery charges and a missed delivery cost far more than a planned workshop slot. Seasonal planning relied on instinct, with no way to quantify which lanes would tighten or by how much.

Underneath every one of these problems sat a data engineering gap. Telematics arrived as raw position pings with no job context, so a vehicle's movement could not be tied to the consignment it was carrying. Without that link, there was no reliable history of how specific lanes behaved, and no basis for training a predictive model. Because the feed also relates to identifiable drivers, any reuse of that data needed to respect UK GDPR expectations around purpose and access.

The human challenge was equally real. The planners had seen confident software before and did not trust a model over their own judgement. Any system that produced unexplained numbers, or that tried to automate decisions they were accountable for, would simply be ignored, so adoption had to be designed in from the start rather than hoped for after launch.

PapaSiddhi began with data engineering rather than modelling. Working through the telematics provider's API, the team normalised the raw feed into a job-aware event stream in PostgreSQL, attributing every position ping to the consignment it belonged to. That job-level history is what made every later prediction possible. A Node.js layer handles ingestion and the operational API, while the predictive work runs as a separate Python service.

Three models sit on that foundation. Delivery ETAs are predicted from historical journey times on comparable lanes, time of day and live traffic from Google Maps Platform, so the system learns that a lane behaves differently at different hours rather than relying on a static distance calculation. Maintenance prediction combines engine hours, load profile and fault codes to flag vehicles whose usage suggests attention before the fixed interval falls due. Demand forecasting projects capacity needs by lane and period, allowing subcontractor cover to be planned ahead.

Everything surfaces in a React dashboard built for the traffic office, where each prediction displays its supporting history and a confidence indicator. Outputs were kept deliberately advisory, with planners free to override any recommendation. The platform runs on AWS with the model service isolated from the web application, so a long training run never slows the operational dashboard. A small dedicated team delivered the work iteratively alongside the planners who would use it.