5-minute interview with Maria Vechtomova

“But I think there is increasingly more understanding of how much value MLOps brings […] Now we are starting to see actual business value. If a model is down you lose money, and for any application you need to have processes in place to be able to track it.”

Meet Maria Vechtomova, MLOps tech lead at Ahold Delhaize. Maria has been working in MLOps before it even was a thing making her a true MLOps pro. Together with a former colleague she writes and shares content for Marvelous MLOps, a newsletter and content channel aimed at helping ML Engineers and Enthusiasts learn more about MLOps.

Tell us a little bit about yourself 

Maria: I'm both a Machine Learning Engineer and MLOps Engineer. I've been in the field for almost 10 years, I started as a Data Analyst, then moved into Data Science and Machine learning Engineering. I've been doing MLOps for the last seven years, so before it became a thing. 

I have built several MLOps platforms from scratch using many different tools, so it is really a passion of mine. Currently I'm working in Ahold Delhaize’s central organization. So we have 19 different brands, mostly food retailers, and it's one of the largest food retail companies in the world. All of the brands basically have the same needs, they want to have the same types of ML models in production and also the same tool stacks. It doesn't make any sense to reinvent the wheel for every brand so I develop and adapt the framework. In large, I act as a product owner and tech lead,  but I also still do some developing myself because I love coding.

I also write content for Marvelous MLOps. I’ve always liked knowledge sharing, for example I enjoy presenting at conferences. At some point I was creating quite a lot of content internally for a use case together with my former colleague Bashak and we really enjoyed that. We wanted to start sharing our MLOps  knowledge, so we started putting up content on social media and through a newsletter called Marvelous MLOps. It started with us putting up some articles and writing some posts and it kind of went from there. We got a lot of positive reactions and it's been a great journey so far, but at the moment it's just a side thing.

How did you get into the ML field?

Maria: I started by studying economics and econometrics at University but I didn't really like it. The only thing I liked about it was the data. Naturally I started looking for jobs related to data. Data analytics was a thing back then, so I was applying for multiple jobs in that area and I chose to work at KPN, a telecom company in the Netherlands. I started there as a Data Analyst, but then I advanced to another department that was called Advanced Analytics who were doing more Data Science. 

In that department we did things in Python and built our first APIs, but back then no one could put those in production. It was kind of natural to transition into Machine Learning Engineering because you had to figure it out on your own, so we ended up creating KPN’s ML framework. It was essentially an experiment tracking system and model registry kind of thing. A lot of functionalities that you see today in these kinds of tools we built ourselves back then, at the time this was pretty forward-thinking. We were also using CI/CD, an orchestration system and then we moved to Kubernetes where we were deploying jobs. It was all an abstraction layer for Data Scientists, they didn't have to do much on their own and it was all very easy to use. They could just take our cookie cutter template with pipelines and everything was set up for them. With this framework we saved a lot of time, because before if you wanted to bring anything to production it took forever. So this philosophy of giving freedom to and enabling Data Scientists, that's what MLOps is all about.

Why is MLOps such an important field?

Maria: MLOps and LLMOps are both important for organizations. Most of the models that you have in organizations are really pretty simple. It's usually something like demand forecasting, recommender systems, classification problems, churn prediction and so on. I would say that more than 90% of everything that companies do is quite basically machine learning. You need to have good systems in place to track how your models are performing in order to deploy it more easily, to make changes more easily and to roll back. Some companies might have those practices in place but are still not doing it right. I think no one is actually really doing it right. But I think there is increasingly more understanding of how much value MLOps brings. I'm sure people have seen models going down for a week and no one notices it. Now we are starting to see actual business value. If a model is down you lose money, and for any application you need to have processes in place to be able to track it.

Also in regards to all the regulations that are coming. The governance aspect is crucial and that's why model registries, feature stores and other tools that are used in MLOps are really useful here. For any model deployment you need to track what code, data, infrastructure and environment was used to produce a result. You also need to know it for any run that you have and you want to roll back. I don't think most ML systems are built with this in mind at the moment.

Do you have any interesting resources to recommend? 

Maria: Through Marvelous MLOps we constantly recommend other learning resources and we’ve created a MLOps roadmap where we name several useful resources. But if I were to just mention some names, I think Paul Iusztin and Paul Labarta Bajo are amazing content creators. We also have Chip Huyen, a bigger content creator, who wrote a great book on ML System Design and has an amazing blog where she goes a lot into LLMs. The MLOps community is also a great resource!

‍

Listen to this and other episodes on Spotify: