How can a recommendation engine predict more reliably what a customer might need, but has not yet bought?
For a wholesaler who sells food to restaurants, we at Data Reply developed a new kind of recommendation engine. Normally, applications like these use data exclusively from the internal warehouse in order to generate product recommendations on this basis for the customers. In addition, our special recommendation engine also uses information from the Internet and in this way facilitates more tailor-made recommendations.
A web crawler looks for general information about restaurants, such as their menus, for our special recommendation engine. At the same time, it also collects data that reveal what was purchased from other wholesalers for the restaurant. This special database has the potential to raise product recommendations to a new level. But only the following measures and MLOps techniques enabled us to exploit this potential.
Training with Kubeflow
We used Kubeflow to train the machine learning model underlying this special recommendation engine. This framework allowed us to orchestrate the machine learning pipeline and consequently ensure an especially efficient project course.
Continuous tests thanks to CI/CD setup
In order to be able to guarantee the quality of our machine learning product, we continuously monitored its performance and accuracy and checked these based on relevant business KPIs. We used a CI/CD setup to do this.
Provision as REST API with Seldon
The platform Seldon enables us to provide the machine learning model simply as REST API and in this way to create a scalable interface for the real application of the model.
Continuous retraining of the model
We use the feedback of the food wholesaler's customers to continuously improve the recommendation engine. We repeatedly carry out retraining of the machine learning model on this basis.
Our new recommendation engine has many advantages for food wholesalers: the better tailored recommendations do more than just boost customer satisfaction and as a result sales. They simultaneously also permit more targeted marketing measures, thus lowering advertising costs. Thanks to the knowledge of what was bought for restaurants in other shops, the wholesaler was also able to expand his product range.
All of these advantages convinced the food wholesaler: the recommendation engine is currently in use in Germany and Poland. However, the rollout in Croatia, Portugal, Austria, Hungary, Romania, France, Italy, Spain and the Netherlands will follow soon.
As part of the Reply group, Data Reply offers a wide range of services that support customers in becoming data driven. We operate in various industries and business areas and work intensively with our customers so that they can achieve meaningful results through the effective use of data. Data Reply offers many years of experience in transformation projects to achieve “data-driven companies”. We focus on the development of data platforms, machine learning solutions and streaming applications - automated, efficient, and scalable - without making any compromises in IT security.