DHL: DATABRICKS MACHINE LEARNING PLATFORM

The situation

DHL wanted to move beyond reactive logistics management towards predictive operations across its European distribution network. Demand forecasting, route optimisation and anomaly detection were the target use cases. Its data science teams needed to develop models, put them into production and monitor them there.

We built a machine learning platform fully in Databricks on Azure. The scope covered the model lifecycle rather than training alone: feature engineering, validation, deployment, real-time inference and monitoring.

Key decisions

We used Databricks Feature Store for feature engineering at scale. This addressed the preparation of model inputs alongside the work of training and serving models. Feature engineering was part of the platform design, not a separate concern left outside it.

We built automated training, validation and deployment pipelines using Azure DevOps and MLflow. The delivery cycle was a concrete constraint: model deployment took 6 weeks. Automating that path reduced it to 3 days, giving the data science teams a shorter route from model development to production.

We implemented real-time inference endpoints with auto-scaling. The platform needed to serve predictions as well as support model development. In production, 12 ML models served predictions for 500K+ daily shipments.

We also designed monitoring for model drift, with automated retraining triggers. That extended the operational scope beyond deployment to the behaviour of models once they were running.

What was hard

The engineering challenge covered several connected responsibilities. Feature engineering had to work at scale; model delivery needed training and validation; inference needed to serve predictions; and monitoring needed to detect drift. Addressing only one of these would not have covered DHL's stated need to develop, deploy and monitor models.

We handled those responsibilities within the Databricks platform, supported by automated delivery pipelines. The results included a 23% improvement in demand forecasting accuracy, reducing overstock and understock events, as well as the shorter deployment cycle.

Lessons for similar teams

Treat model delivery as part of the platform, not as work that begins after training. DHL's deployment improvement came through automated MLOps pipelines spanning training, validation and deployment.

Plan for inference and monitoring alongside development. Real-time endpoints, auto-scaling and drift detection each address a different operational requirement.

Finally, distinguish model quality from delivery performance. Forecasting accuracy and deployment time measure different things. In this engagement, both provided concrete evidence of what the platform delivered.