Ahold Delhaize’s Machine Learning Operations: Gaining Insights.

April 5, 2023

Ahold Delhaize’s Machine Learning Operations: Gaining Insights

April 3, 2023 – Machine learning operations, or MLOps as it is more widely known, is a team effort frequently made up of data scientists, devops engineers, and IT. It is a fundamental component of machine learning engineering and a crucial and practical method for the development and effectiveness of AI (Artificial Intelligence) solutions. For a global food retail technology business like Ahold Delhaize, it is crucial.
It can be difficult for many businesses to put machine learning models into production, so networking with other businesses would be beneficial. Ahold Delhaize held its first MLOps conference on March 23–24 in order to exchange knowledge across sectors and aid IT specialists in coming up with fresh approaches to persistent issues.
The conference, which was held at our global support office in Zaandam, the Netherlands, was a fantastic event that brought together more than 100 people who shared the same interests and where inspiring tech leaders from top-tier Dutch businesses discussed the problems and solutions they are working on. Three major subjects were covered by the conference and speakers:

What tools can be used and how infrastructure for machine learning initiatives should be organised
How to set up teams to expedite the time it takes for machine learning goods to reach the market
tackling difficult machine learning tasks
Important quotes from the presenters
The following insights from MLOps projects were shared during the event by speakers from (but not limited to) ING, Lely, Schiphol, codebeez.nl, Albert Heijn, bol.com, and Ahold Delhaize.

Joshua Lee, ING’s Tech Lead
Joshua explained why implementing MLOps in an organisation can be difficult. His main conclusion was that there is no one “right” MLOps instrument for every task. All of the equipment required for MLOps already exists in a big organisation. It is crucial to put consumers and data scientists first rather than fancy tooling.

Platform engineers Anastasija Efremovska and Quiran Storey from bol.com
Anastasija told their tale of how self-service developer tools came to be. They revised the prior argument, which targeted java microservice developers as the primary audience, and expanded it to include other user groups, including the data science community and programmers creating bespoke infrastructure. It significantly improved the degree to which all user groups were genuinely autonomous.

Albert Heijn data scientist Guus Verstegen
In order to cut waste and improve on-shelf availability, Guus discussed how he is developing a daily demand forecast system for more than 1,000 stores and more than 20,000 products. Key takeaway: For Albert Heijn to create and support such a critical data science product, the right tooling selection, an autonomous working style, shared knowledge, and team responsibility are all essential.

Ahold Delhaize’s Mariia Vechtomova, MLOps Transformation Manager, European Business Services
Mariia talked about creating an MLOps framework that can assist data science teams in drastically reducing the time it takes to bring machine learning goods to market. The key lesson is to create MLOps frameworks pragmatically, start small, and demonstrate value as soon as possible.

Tim Rietveld is a Lely Machine Learning Engineer.
Tim concentrated on making machine learning models accessible to peripheral devices. He talked about the difficulties that come with the various toolkits that machine learning and embedded programmers use, as well as the need to increase team autonomy. As more organisations need machine learning models on edge devices, this subject is becoming more and more important.

Albert Heijn machine learning engineers Paolo Radaelli and Mariska van Willigen
Mariska discussed Albert Heijn’s dynamic markdowns, a key machine learning tool that helps the company achieve one of its main objectives: waste reduction. Mariska and Paolo discussed the shortcomings of the present design and how the new architecture addresses them. It can be challenging to change the design of a product that is already in use; for this reason, it’s crucial to have open lines of communication both within the team and with the stakeholders. Also, keep your team small and your project moving quickly.

Data engineer Justin van Dongen from Schiphol
Justin spoke about the reusable framework for data science goods. Data will play a significant part in helping Schiphol achieve one of its goals, which is autonomy. This leads to an ever-expanding area of AI where the need for control increases. Justin described how he used MLOps principles to help make their system more maintainable while maintaining flexibility for the Data Scientists.

Arne Muller, European Business Services Machine Learning Engineer at Ahold Delhaize
Arne provided information on how to better implement a cross-selling model that uses an MLOps framework to recommend products to customers at the checkout page for several Ahold Delhaize brands, including Mega Image, Maxi, Alfa Beta, and Delhaize. Deployment to a different brand can now be accomplished within a month instead of the prior 9 months thanks to the MLOps framework.

Python AI Engineer Luuk van der Velden from codebeez.nl
Luuk spoke about his work at NS and how they used stream processing to update timetables and anonymize customer data while setting up lambda architecture for real-time train crowding forecasts. The main lesson is that you should carefully select your architecture while considering costs and development complexity into account.