5-minute interview Jiri Steuer

“For serious discussions surrounding machine learning [...] you need proper tooling, procedures and equipment. For me that is MLOps together with a feature store, there is no other way to do it.”

In this episode of 5-minute interviews we meet Data Architect Jiri Steuer. This week we talk about why he thinks MLOps and feature stores are indispensable when it comes to building machine learning applications.

Tell us a little bit about yourself

Jiri: My name is Jiri Steuer and I am from EmbedIT that has a tight relation to the company Home Credit International. We are primarily focusing on delivering complex IT services for different industries and I am focusing on enterprise solution architecture for data and applications. From a history point of view, I have 15 years of experience in the finance industry and four years in the tech industry. I started with C, C++ and machine learning at University where I focused on analysis of writing text, combination tasks with focus on genetic algorithms and basic usage of neural networks. 

How did you get into the ML/MLOps field?

Jiri: I started to focus on MLOps and feature stores around 4 years ago, because we focused on time to market improvement for the delivery of machine learning and AI together with data strategy. Data-driven companies require a solid data strategy and it is an important enabler for data monetization. And this was the reason why we focus on this important area. The main use case for us is definitely the financial industry. But we are focusing on a lot of broad problems with customers and based on this you can see the application of machine learning in many areas. Not only finance but telecom, TV,  enterprise and so on.

What attracts you to the field of ML/MLOps?

Jiri: I think it is more clear than ever today that machine learning and AI can improve efficiency in many areas. I see a strong parallel between standard data development and machine learning development. For clarification, let me give you a small example. In data development, you talk about standard terms such as data layer historization, data model description, data quality classification, master data management. When you have to start a serious discussion about machine learning, you will see the same or similar needs. You can also see the layers with input data, calculation of the predictor, outputs from the machine learning model. You can also see the proper description of contents. You also have to cover deeper quality checks with domain knowledge for predictor calculations, also to the topic master data management because you need to cover or create a new predictor as the new asset. For this reason, ML and AI together with MLOps and feature stores is, for me, an extension of standard data development and a natural way for data monetization. It is an easy formula for me.

I'm a fan of proper work, but I can also imagine the freestyle development of machine learning without using a feature store. But for serious discussion and serious functions and capabilities (surrounding machine learning), I see this clear way for how to do proper descriptions, how to improve the quality of everything in time etc., and for that you need proper tooling, procedures and equipment. For me that is MLOps together with a feature store, there is no other way to do it.

Do you have any interesting resources to recommend?

Jiri: To be honest, I can’t say only one source. But as an example, I think Gartner’s studies are very useful. Based on practical experience, look at the leaders of the MLOps and feature stores industry and you can see a lot of companies to learn from, for example, Iguazio, Hopsworks, Databricks and Cloudera, which are all focusing on machine learning. But it’s hard really to give any good definite resources.

‍

Listen to the full episode: