flaschenpost SE is one of Germany's leading online supermarkets.. The Münster-based company, founded in 2016, can deliver your entire weekly shop to your home or office in just 120 minutes. Customers in more than 200 cities and almost all metropolitan areas in Germany benefit from flaschenpost's logistical and technological expertise. Around 20,000 employees now work for the company, which ships more than 10 million orders annually.
The project at a glance
- Creation of a resilient data product with a high degree of reproducibility
- Significant reduction in the cycle time from model development to evaluation in live operation
- Significant reduction in IT infrastructure costs
- Unreserved customer recommendation
Initial situation
When it comes to the delivery of beverages and other food products to your doorstep, the ability to process orders quickly gives you that crucial competitive advantage. No more than 120 minutes should pass between an order being placed in the webshop and it being delivered to your front door.
flaschenpost uses machine learning to organize driver logistics in order to create forecasts for its logistical processes and supply chains. This involves processing and analyzing large volumes of data in order to identify patterns and trends that help predict driver availability and estimate the duration of work steps. The more precise the estimates are, the more accurate the route planning results will be for the drivers.
Failure to predict the number of drivers needed at a particular time often results in under- or overutilization of driver capacity. Underutilization can lead to idle waiting times and possibly rejected orders, while an overload can cause delays and cancellations of orders that can no longer be filled. Being able to predict these situations accurately is therefore a direct route to greater customer satisfaction and increased sales.
Solution
A machine learning system was already in place at the start of the project, which was accessible as a platform implemented in Microsoft Azure. It received more than 10,000 forecasting queries a day.
Team structure
Project milestones
Technical design
Team structure
A joint team consisting of developers from flaschenpost and codecentric was formed for the duration of the project. The team comprised a mix of experienced and young data scientists and data engineers using agile methods such as scrum to continuously improve and operationalize the algorithms.
Project milestones
An important milestone was to expand the forecasting portfolio so that further steps in the supply chain could be optimized for machine learning. In addition, various solutions for the automation of complex data and machine learning pipelines were investigated and implemented. The aim was to reduce the workload associated with administrative and operational tasks. It was also necessary to display the metrics of different algorithms in a clear way and to be able to compare them. This made it possible to quickly identify the best model and to systematically investigate which data leads to an improvement in an AI model.
Technical design
The settings that need to be defined before training a model and that influence its performance were automated. Monitoring and reporting were enhanced in a meaningful way so that the team has an even better overview of the performance of the forecasting portfolio.
The main programming language was Python, which, in addition to being very well integrated into Kubernetes and Azure, also supports all relevant machine learning and data engineering frameworks. The technology stack also included tools such as Kubernetes, MLFlow, HyperOpt, FastAPI, and streamlit.
Result
flaschenpost SE's optimized and enhanced logistics forecasting portfolio is now running smoothly and can be easily extended. The solution makes it possible to systematically analyze the data and thus improve the AI model. Automation, tracking with MLFlow, and a dashboard that displays the metrics of each algorithm significantly reduced the cycle time from model development to evaluation in live operations.
Developers at flaschenpost can now transparently view the relevant data, training parameters, and model utilization. What's more, the data can be easily adapted. The models developed in this way enable flaschenpost to better predict and plan driver logistics before, during, and after a delivery round. Automation has both reduced IT infrastructure costs and increased data analysis speed.
Dr. Jens Rauch
Product Owner Team Data Science Logistics Forecasting
codecentric provided us with excellent support during a major overhaul of our machine learning models for route planning. We can now extend, manage, and deploy our models more easily, and better evaluate and reproduce experiments. We also worked with codecentric to significantly improve the accuracy of our forecasts. I can wholeheartedly recommend codecentric.
Dr. Jens Rauch
Product Owner Team Data Science Logistics Forecasting
Any questions about the project?
Would you like to improve your processes in a data-driven manner? Or would you like to revitalize existing data projects?
IT Consultant / MLOps Engineer
Further projects of codecentric AG
Find out about further successful projects that we have completed with our customers. Perhaps you will find inspiration for a use case in your company.
Patrick Soschinski
IT Consultant / MLOps Engineer