Production-Ready Machine Learning Framework
Name
Kseniia Leshchenko
Abstract
Machine Learning (ML) has quickly progressed from a strictly academic field to a discipline that is bringing real value and competitive advantage to the organizations that leverage it. ML allows to automate decision making in the areas where the traditional rule based systems fall short due to the complexity of defining or maintaining the underlying rules. Developing and maintaining production-ready ML systems has, however, proven to need a lot of know-how, time, and money. Moreover, the field of production-ready machine learning is still emerging and the tools and best practices along with it, leaving a gap between the desire of having a solution and actually being able to have one that works reliably over the time.
This work proposes a general Machine Learning Framework, which introduces
production-ready practices for designing ML systems. Furthermore, the framework also comes with application and deployment level code templates that allow to set up a new ML solution utilizing arbitrary ML methods and models. The framework simplifies the process of building an ML system with monitoring, Quality Assurance, and customization principles in mind.
The framework is successfully applied in a major system in production that has been serving millions of requests and has been operational with virtually no downtime for over a year, while ever improving and meeting the scalability requirements. Moreover, the framework has altered the mindset of local MLOps engineers, has made an impact to AI solutions marketing, and has reduced the estimated budget for new AI proof of concept projects.
This work proposes a general Machine Learning Framework, which introduces
production-ready practices for designing ML systems. Furthermore, the framework also comes with application and deployment level code templates that allow to set up a new ML solution utilizing arbitrary ML methods and models. The framework simplifies the process of building an ML system with monitoring, Quality Assurance, and customization principles in mind.
The framework is successfully applied in a major system in production that has been serving millions of requests and has been operational with virtually no downtime for over a year, while ever improving and meeting the scalability requirements. Moreover, the framework has altered the mindset of local MLOps engineers, has made an impact to AI solutions marketing, and has reduced the estimated budget for new AI proof of concept projects.
Graduation Thesis language
English
Graduation Thesis type
Master - Software Engineering
Supervisor(s)
Karl-Oskar Masing, Raivo Kolde
Defence year
2022