Go fast by going well

Industrialising the use of machine learning and AI at Novo Nordisk. How we’re building the necessary infrastructure to move experiments into production, safely and at speed.

The first industrial revolution in 1760 lasted 80 years, with mechanised production equipment and steam engines boosting efficiency. Today, businesses stand on the brink of the AI and machine learning revolution; to apply the next generation of tech, to transform industry, research, and production at a breathtaking pace. At Novo Nordisk, a dedicated machine learning operations team are ensuring the swift transition of AI projects from experimental to operational phase. Thomas Henckel, Lead ML Engineer in this team, explains their role.

Checkmarks for good practice

Whereas the first industrial revolution greatly improved efficiency, there was little consideration for safety and health. Today, these are, of course, key issues in most industries, and especially in the pharmaceutical industry, where the main product is life-saving medicine. To ensure the highest quality of their product, Novo Nordisk complies with stringent practices set by the authorities and the company itself. These are known commonly as GxPs, meaning Good something Practice, where the x can be substituted by M for manufacturing, C for clinical – and the list goes on. Ensuring adherence to these practices is now something that machine learning can support and make more efficient.

Identifiers and facilitators

But, of course, it’s much more than compliance, although that is a key component. It’s also about facilitating the work of researchers; bringing their ideas from experiment to operation, and to scale their work within a larger company framework. In this process, the AI and Analytics Centre of Excellence in Novo Nordisk are both scouts and facilitators. They identify machine learning and Generative AI projects that are mature enough, in terms of a high probability to create value, and then they go in with technical support to help the researchers cross the challenges that they might be facing when moving from experiment to operational mode – guiding them at every turn.

Complementary skills

Technically, this is done by introducing good software practices, centring the collaboration around a source repository, and introducing DevOps practices to ensure traceability and scalability. With the number of possibilities and flexibility in today's data science platforms, the MLOps team collaborates with the researchers in configuring the platform, so it suits the needs of the project and the researchers. They do this to take complexity out of the process, making it easier for the data scientists to focus on performing their experiments. In this sense, the team focuses on enablement. They bring the tools, the infrastructure, and find the missing personnel to support bringing projects into operation.

Skill wise, the team is built to have a solid understanding of how to connect data science, machine learning, and software engineering. This complementary understanding is the foundation for addressing the challenges that researchers may have. It's crucial for transforming these challenges into operational processes that can be scaled across the organisation.

Let’s do it well

And that’s where we are today. At the cusp of another industrial revolution, where machine learning can massively change production flow and processes. But, although fast is a keyword, nothing happens without considering quality management frameworks, which are the cornerstone of every production process in today’s pharmaceutical industry.

“In the MLOps team we are proudly sharing our department’s quality mindset that is often spoken out using Robert C. Martin quote “The only way to go fast is to go well”. That's why we come in with engineering capabilities, the security, and the compliance that makes sure that your Machine Learning project doesn't strand as a garage ornament or a showcase, but actually gets out to make real value.”

Thomas Henckel, Lead ML Engineer, MLOps team.