Forecasting and inventory work, scoped small enough to ship

A cockpit that plans the network and reacts when it breaks

Customer demand lived in one system, returned supply in another, transport costs in a third. Planners could say what had happened, usually a week late, and never what it had cost. Rebalancing the network was an annual exercise done in spreadsheets, and every sudden event — a storm, a line down, no trucks — was handled by a chain of phone calls.

We built a network optimisation platform that models customer demand together with the probability of supply being returned to each plant, prices both against real transport costs, and turns the result into rebalancing recommendations: which plant should serve which demand, and what it costs to get that wrong. On top of it sits a full supply chain cockpit — one set of numbers for planning, for operations and for the people who sign the invoice.

• Return probability estimated per plant rather than assumed, because the flow back into the network is what quietly decides real capacity.
• Every rebalancing recommendation carries its transport cost, so the trade-off is visible at the moment of the decision instead of in next month's report.
• An operational layer that learns every day from which recommendations planners accept and which they reject, so it stops proposing what this team never does.
• Sudden events — weather, production shortfalls, transport capacity — arrive as proposed actions with their cost attached, not as an alert somebody has to interpret.
• Built on Databricks, so the same lakehouse feeds the optimisation, the cockpit and the models without a second copy of the truth.

What transport should cost, and who will actually take the load

Rates were negotiated against last year's rates. There was no way to tell whether a lane was expensive because the market was expensive or because nobody had looked at it in three years. And when a carrier rejected a load at the last minute, covering it was a scramble against the clock with no idea what the right price was.

We benchmarked the transport operation against the industry lane by lane, and built a recommendation engine on top of it that reacts to last-minute rejections: when a carrier drops a load, it proposes which carriers to go to next and at what price, ranked by who has actually run that lane, how reliably, and at what cost.

• Each lane compared against the market rather than against its own history, so the conversation with a carrier starts from evidence instead of from last year's number.
• Rejections treated as pricing information, not as incidents: a carrier that keeps turning down the same lane is telling you your price is wrong before your spend does.
• A proposed price attached to every carrier suggestion, so whoever is covering the load at six in the evening is not negotiating blind.
• The same lakehouse behind both the benchmark and the recommendations, so what the analysis says and what the tool proposes cannot drift apart.