blog Article

4 Must-Do’s When Starting Your MLOps Journey

Author: Jan Wuzyk ,

Every day, many companies incorporate machine learning (ML) into their software or application projects. And company leaders have begun to realize that building ML models requires a very different approach than traditional software development.

As the scale and complexity of each AI project grow, the greater is the overhead when it comes to model development and deployment. One way to reduce the complexity and overhead of AI projects is to adopt MLOps. It helps you better control the process and keep producing results quickly as the project grows.

What is MLOps?

MLOps is a set of principles, practices, and tools that enable companies to quickly and reliably deploy software powered by machine learning. By applying MLOps practices, companies can simplify the management of ML models, making them easier to deploy in large-scale production environments. MLOps builds on top of DevOps, attempting to solve the challenges of traditional software development while also considering the challenges ML models present.

The Benefits of MLOps

If you build AI-powered products, then you should consider implementing MLOps (if you haven’t already) because it provides many benefits, including:

  • Involves automating processes to speed up the creation and deployment of ML models, which means AI-driven applications and products get to market faster.
  • Encourages the monitoring and evaluation of models during production, ensuring that models always perform well and avoid model drift.
  • Facilitates collaboration among different teams with various skill sets, which helps improve the outcomes of AI projects.

The longer you apply MLOps practices to your AI projects, the more benefits you will see.

Start MLOps with these four steps

If you want to implement and gain the benefits of MLOps, you should start with these four steps:


Four steps to start MLOps1) Encourage collaboration among teams.

You can overcome any challenge in ML-powered software development when teams have access to those with relevant experience. However, you’ll often find that the people with relevant expertise rarely work directly with each other. You should encourage collaboration among everyone involved in the entire lifecycle of a machine learning solution, perhaps creating cross-functional teams. At Radix, we have found that collaboration through cross-functional teams massively speeds up the rate at which we deliver solutions and makes tasks easier for everyone involved. 

2)  Make sure teams have the right tools. 

You can’t build high-performing models or successful AI-driven applications if teams don’t have access to the right tools. For example, training models require a lot of compute resources and hardware. So, you need to make sure data scientists have easy access to powerful computing resources, such as Microsoft Azure or Amazon AWS. Teams also need to keep track of experiments and make sure they are reproducible — failing to do these things will lead to poor model performance or regressing. You can track experiments with tools like MLFlow or wandb. And to ensure reproducibility, you can use tools like git, dvc, or poetry.

3) Automate as much as possible as early as possible. 

For decades, software engineering teams have been grappling with the problem of how to deploy products quickly and reliably. To solve this problem, these teams came up with the concept of Continuous Integration/ Continuous Deployment (CI/CD), which involves introducing automation into various processes to speed up app development. You should embrace automation, adding it to as many steps of your training and deployment pipelines as possible, supported by automated tests. With automation, you can deploy models faster and as early as possible. You can deliver tangible results early, so that model deployment and training grow together, which helps prevent challenges in deploying your model further down the line.

4) Monitor and retrain your models continuously.

Machine learning models are not as stable as traditional software and thus require a lot more care once deployed. It’s common that model performance decays over time as the incoming data drifts away from what the model was trained on. The first step to mitigating this problem is monitoring the model’s predictions. In addition, you can use data seen in production to retrain your model. Monitoring and retraining your models will help you keep your model performing well. Again, as with deployment, this process should be as automated as possible.

MLOps in Practice

At Radix, we pride ourselves on creating high-quality AI solutions quickly — and applying MLOps practices to all our projects is a core component of how we achieve this. We all strive to achieve the moniker of “full-stack machine learning engineer,” which means that every engineer fully understands the ML solution development process — from ideation to deployment and beyond — while still having their own specializations. This approach ensures a smooth transition from the model development phase to deployment, allowing us to deploy models rapidly.

We also strive to follow best coding practices, as exemplified by Poetry Cookiecutter, an open-source Cookiecutter template for scaffolding Python packages and apps created by our CTO Laurent Sorber. With Poetry Cookiecutter, you can quickly create and maintain Python code projects, using Cookiecutter templates as the base. 

We have experience across many project types and scales — from deploying proof of concepts to large-scale projects requiring complex cloud deployments to run experiments. We also help companies apply practices to their own projects, recently a project for a Brussels Airport Company.

Don’t Miss Our Upcoming MLOps Webinar!

On June 15, 2022, at 11:00 am CST, Radix and Brussels Airport Company will hold a webinar to discuss MLOps. Join Radix’s Brecht Coghe and Xavier Goás Aguililla and  Brussels Airport Company’s Thibault Verhoeven in a discussion that will include how MLOps helps Brussels Airport reproduce and scale their AI models. You will also have a chance to ask these experts your questions. 

Don’t miss this chance to learn more about MLOps!

Register for the webinar below.


Jan Wuzyk
About The Author

Jan Wuzyk

Jan is a Machine Learning Engineer, originally from South Africa. He has a master’s in Mathematics from the University of Bonn in Germany. Jan loves Machine Learning because it melds mathematics and programming with impressive and useful results. He is excited to keep exploring the world of Al and working on rewarding projects.

About The Author