From AI Experiments to Production with MLOps

Elevating AI to Production: The Rise of MLOps

AI is transitioning from isolated experiments to robust, large-scale production solutions. A key element of this transition is the increased adoption of MLOps, the AI equivalent of DevOps. MLOps provides companies with a documented, optimised, and secure process, resulting in significant business gains.

The AI Boom and the Need for Structure

While interest in AI surged in 2023, thanks to OpenAI’s ChatGPT, companies are only now starting to implement AI strategically to achieve sustainable growth. A McKinsey report from May 30, 2024, reveals that global AI adoption has risen to 72%, up from around 50% in the previous five years. Over 50% of surveyed companies now use AI in multiple functions, compared to under 30% in 2023.

From Scattered Experiments to MLOps

The approach to AI has evolved. Initial limited experiments have paved the way for large-scale AI implementation to fully leverage its potential. This means AI must be integrated into entire production solutions. Consequently, AI development requires the same systematic approach used in software development, including automated build, testing, and monitoring.

This methodology is called MLOps, which streamlines processes, shifting from manual AI experiments to production-ready, maintained, and monitored solutions.

The Benefits of MLOps

Currently, the output produced by data scientists often goes into production manually, increasing the risk of errors and limiting monitoring capabilities. A streamlined process ensures an efficient and robust product with reliable models and the necessary response times and uptimes for customer-facing applications.

7 Key Principles for MLOps

• Use version control for all code and data.
• Utilise multiple environments (development, test, and production).
• Maintain infrastructure as code (IaC) under version control.
• Ensure full traceability of experiments.
• Automate testing for code, data integrity, and model integrity.
• Employ Continuous Integration (CI) and Continuous Delivery (CD).
• Monitor services, models, and data.

Traceability, a crucial aspect of MLOps, ensures transparency and compliance by enabling explanations of model behaviour, the code and data used for training, and comprehensive test reports. This aligns with requirements such as the EU’s AI Act.

Combining Off-the-Shelf Products and In-House Development

Modern AI development often involves a blend of software code, APIs, custom-built components, and pre-built models from cloud providers or open-source communities. Microsoft, with its co-pilots and Azure services, offers both ready-made apps and building blocks for creating complete solutions.

This approach necessitates careful orchestration of versions, dependencies, and automated testing to ensure overall stability.

MLOps: A Journey of Continuous Improvement

Implementing MLOps can be complex. Beyond technical skills and robust processes, MLOps must accommodate the inherent uncertainty of AI, allowing for experimentation with data and models.

Therefore, it’s advisable to introduce MLOps incrementally, gradually increasing maturity levels. Microsoft’s data suggests five levels of maturity.

Practical Applications of MLOps

Here are examples of how companies leverage MLOps to optimise their models:

• A retail company uses MLOps to analyse data and identify optimal store locations.
• An online business employs MLOps to understand user behaviour and offer relevant products.
• A streaming service utilises MLOps to process user data and personalise content.
• A medical research unit uses MLOps to streamline and automate the analysis of large datasets.

In all these cases, MLOps ensures robust AI models in production, continuously monitored to perform their designated tasks effectively.

Devoteam Cloud Enabler for MLOps

Devoteam offers a method and accelerator for implementing MLOps in Azure, prioritising time-to-market and solution security. Their accelerator utilises the latest Azure components, including Azure Machine Learning, Azure OpenAI, and Azure Kubernetes Service. This method, called Cloud Enabler for MLOps, includes methodology, documentation, infrastructure, MLOps automation, monitoring, and end-to-end pipeline building jobs. It supports both simple machine learning models and advanced scenarios like fine-tuning large language models (LLMs). The resulting product is typically a container with an API, hostable in the Azure Cloud or on-premise.

Devoteam’s approach focuses on helping customers quickly establish an MLOps platform while empowering them with the knowledge and methods for independent operation.