Capral Limited · ASX:CAA · Manufacturing
Predicting when an order will arrive
Delivered
The question
Capral’s customers ask a simple thing: when will my order arrive? Answering it well is a core part of the customer service operation, and the existing answer came from a rule-based method that could not account for the variability of a national extrusion and logistics network.
Order data lives in SAP ECC. What happens on the presses lives in plant-level manufacturing execution systems — AMS, FMS and Ignition — across plants that are not all on the same platform. A useful prediction has to join the two.
Australia’s largest manufacturer and distributor of aluminium extrusion products, operating across multiple plants nationally.
How it was built
Four phases, each ending in a written go/no-go. The first phase — discovery, and a read on whether the data would support a model — was delivered at no cost.
Ingest. SAP ECC extracts and MES telemetry land in Azure, then move through a medallion architecture on Databricks.
Engineer features. Under point-in-time discipline: at any moment in the training history, the model sees only what each system knew at that moment.
Measure the incumbent. The existing rule-based method was replicated over historical orders to set the baseline.
Serve it. Customer service queries the prediction in natural language, through a conversational interface.
Feature design
Three tiers, starting with what the plant teams already knew.
First, the bottlenecks the plant teams identified when we walked the floor at Penrith and Bremer Park. Then what the data showed on inspection. Then systematic search.
In that order, the model starts from what the business already understands about its own process, including the parts that never appear cleanly in an export.
What success meant
The evaluation dataset, the baseline calculation method and the minimum success threshold were agreed in writing before the build began. The result was measured against that threshold.
The SAP extraction patterns were designed for reuse across Capral’s wider Databricks reporting programme, so the engagement also left a reusable integration pattern.
Next
Production deployment, MLOps for retraining and monitoring, and automated SAP ingestion in place of the manual extracts used during the proof of concept.
Technologies