All work
Case N° 02 / 03
Aldoria Logistics · Freight & Supply · 2024

Predictive routing across 14,000 SKUs and 38 European depots.

We replaced Aldoria's static route planner with a forecasting and routing stack that learns from every delivery — and we did it without disrupting a single overnight run.

Plate — 022024
Aldoria Logistics — cover image
Rotterdam · Lyon · ManchesterAldoria Logistics
Client
Aldoria Logistics
Sector
Freight & Supply
Year
2024
Region
Rotterdam · Lyon · Manchester
Duration
14 months · three phases
Team
Three ML engineers, two backend engineers, one ops analyst, one delivery lead
Brief

Aldoria moves time-sensitive consignments across thirty-eight depots in seventeen European countries. The board had set a target of £10 million in annual cost savings within twenty-four months, with no degradation in service-level performance. The existing planner had been built in-house in 2014 and had reached the limit of what spreadsheet logic could do. We were brought in to design and build the replacement.

Challenge

The existing planner was wrong about 18 per cent of the time, which the dispatch teams quietly worked around using twenty years of accumulated tacit knowledge. Any new system had to be measurably better than that workaround — not just better than the planner — and it had to be trusted by people who had every reason to be sceptical. The fleet was 1,400 vehicles across four legal entities, each with its own labour rules. And we could not interrupt overnight delivery, which moved roughly £3.4 million of freight every twenty-four hours.

Approach

How we went about it.

01
Forecasting before routing
Most routing projects start with the optimiser. We started with the demand signal. We built a forecasting layer — gradient-boosted models per SKU per depot per day — and only once that layer was trustworthy did we plug it into the routing engine. Forecast quality, not optimiser cleverness, determines how good a plan is.
02
A digital twin of the network
We rebuilt the network as a simulation environment first. Every depot, every shift pattern, every customs nuance. The new planner ran in the twin against three years of historical daily plans before it touched a real vehicle. That gave the operations directors something they had never had: a way to argue with a routing decision and have the system show its working.
03
Shadow weeks, then sleeve weeks, then full
We rolled out the planner in three layered phases. Six weeks of shadow — the system suggested, the dispatchers ignored, but every disagreement was captured. Eight weeks of 'sleeve' — one corridor at a time, dispatchers had a one-click override. Then full rollout, depot by depot. There was never a cutover. There was never a moment the old system was switched off in anger.
04
Built for the people who use it
Dispatchers do not want a black box. We built an interface that explains every routing decision in two sentences and lets them re-plan a single leg without re-running the optimiser. Adoption was not a training problem; it was a respect problem.
In their words
Twenty years on the desk. I have never trusted a routing system before. I trust this one because it shows me why.
Head of Dispatch, Aldoria Logistics — Rotterdam
Outcomes

What it delivered.

£11.4M
annualised cost saving · year one
+5.4pp
on-time delivery
−71%
customer credits
4.1%
MAE at 14-day forecast
1,400
vehicles in scope
01
£11.4 million annualised cost saving in year one
Audited by Aldoria's finance team and signed off by the external auditors. The board's twenty-four-month target was met in eleven.
02
Service quality moved up, not down
On-time delivery rose from 91.4 per cent to 96.8 per cent. Late-delivery customer credits dropped by 71 per cent in volume.
03
Forecast accuracy that compounds
SKU-level demand forecasting is now within 4.1 per cent mean absolute error at fourteen days out — better than every commercial system Aldoria benchmarked us against, and improving every quarter as the model retrains.
Selected stack
Python · PyTorchXGBoostRayApache AirflowSnowflakeKubernetesMapbox · OSRMTerraform on AWS
A similar problem in your business

If the shape of Aldoria Logistics’s problem rhymes with one of yours, the most useful conversation is rarely an email exchange. We will sit with two of your operators for an hour and tell you whether we can help, whether someone else can help better, or whether the problem is not yet ready to be solved. That conversation is on us.