An efficient infrastructure for KLM’s predictive models

Machine Learning AI Data

Data and self-learning algorithms support KLM’s operational decisions.

Anyone who books a flight ticket with KLM must think about what directly and indirectly influences that booking. From the flight schedules to the amount of water, an airplane must carry, or the number of counters open at the airport, there are dozens of factors an airline must be able to anticipate. Moreover, they are also constantly changing. No surprise, then, that behind the calculation of all those factors lies a massive amount of data. KLM’s Operations Decisions Support is the department responsible for processing all operational information to produce valuable forecasts that form the basis for, for example, work schedules or the number of meals an aircraft must carry. Part of this is done through self-learning algorithms, part of various predictive models that know a range of end users. KLM called in LINKIT to ensure that these algorithms can be used correctly and have this knowledge in-house for the future. 

Machine-Learning Engineer and Solution Architect Alyona led the team at KLM and talked enthusiastically about the assignment. A complex matter can be explained as follows: developing and setting up the structure around the artificial intelligence that is already in use. It means that the LINKIT team, which is part of the team at KLM, ensures, among other things, a stable and continuous data flow for data scientists (who develop the algorithms and therefore need data to experiment and test with). 

In addition, they ensure that the various components required to make the predictive models work in the correct form and order. “The KLM team,” says Alyona, “is well versed. They have the people who develop the models, and there are the people who design the algorithms that support these ‘products’. What they didn’t have was the technical engineering capacity.” That is why LINKIT brought in Data Engineer Alberto, Cloud Engineer Anthony, and Machine-learning Engineer Nastiia. “That is an addition to the team, which is necessary to use the products within KLM correctly.” 

A defined assignment 

Daan van den Oever, Director of the team that makes Machine-Learning models at KLM-ODS, talks about the start of the project. “KLM is working very hard on digitization to improve day-to-day operations. Part of that is the application of AI. We missed seniors for that who did data monitoring and built the cloud environment around it, for example. So, you need people to ensure that the models are trained automatically, monitored automatically, etc.” LINKIT was therefore flown in with a clear package of tasks: the team was responsible for setting up the infrastructure, supervising the cloud transformation, and assisting in recruiting and training junior and medior KLM engineers. It will allow KLM to continue the work in the future with its engineers. “We specifically didn’t want any ‘hands’,” says Daan, “our goal was a well-defined assignment.” 

Get started with the sample 

Before the arrival of LINKIT, KLM worked with one large model that solved various business cases. “The big monster”, Alberto and Alyona call it. Dan laughs. “Gradually, even before the arrival of LINKIT, we already found out that we had to pull this apart. But that, too, was only possible with the right infrastructure. Training, retraining, maintaining and monitoring a model is only possible with the right automation.” The decoupling of the model is, therefore, in full swing. “We have now gone from one to five to eight different models,” says Daan. “Because it only really delivers if you can keep about twenty models in the air with, say, twenty people. Then you can start scaling exponentially.” Because KLM has engineers at its disposal, it is also easier to ‘deploy’ a product. Alyona explains: “By applying engineering knowledge and automating tasks, data scientists can make their model accessible to stakeholders much faster. And suppose something strange happens to the data in a certain place. In that case, we can also provide alerts through this automation, for example, so that a data scientist can immediately look at that separate model.” It makes deployment a lot simpler.” 

From on-premise to off-premise 

The ultimate goal is to move the entire environment to the cloud. “Artificial Intelligence requires a huge amount of computing power, and you want that to be scalable,” says Alyona. A solution in-house, so ‘on premise’, is much less efficient. In the cloud, you are flexible and can take advantage of cloud-native solutions: tools that take full advantage of the scale, resiliency, and elasticity that the cloud offers. However, it is pretty close. “To at least help KLM migrate specific components that are not available on-premise, we already use cloud-native solutions,” says Alyona. Meanwhile, the team works with cloud architects at different levels for the migration. “We are very pleased that the necessary data links have been made to make data traceable and accessible. But moving to the cloud saves time and money, so that’s our next goal,” says Alberto. 

The right tasks with the right people 

LINKIT also brought about the necessary structural reforms on an organizational level. For example, they split the engineering part and the work of the product developers and data scientists, which resulted in three teams instead of one. Alyona: “Engineering is horizontal, in other words: applicable to any product. The work of data scientists and product developers is different for each product. By splitting that, people could work much more efficiently.” At the same time, automation and setting up new workflows also make for a more productive team. “For example, by having the raw data automatically end up in a central location, all data scientists have access to the same data,” Alberto adds. “We can now have datasets generated automatically. It saves time and makes all results transparent and reproducible for everyone and ensures that people enjoy their work because they can focus on what they were hired for.” 

Continue with our own experts 

The most significant added value? According to Daan, this comes from knowledge sharing. “And not only technically, but also process-wise, how you build your teams and how they have to work together or how you give people the right guidance to enable them to grow.” Alyona: “Our goal was to bring more transparency to the process in all areas. That gives confidence and changes how people view the process. There is no more room to experiment.” The technical added value? KLM-ODS will be able to put more models live with a smaller or the same size team in a year. “In this way, we will be able to deliver much more value with the same number of people,” says Daan. In 2023, the KLM team hopes to be able to pull the plug on the old model for good. “The first models will then be in the cloud; they will be automatically monitored and preferably automatically retrained,” says Daan. The LINKIT team will then have left. Their job is done.