Practices for sane MLOps Engineering: Operational AI in a Non-Development-Centric Company

Abstract

Transitioning from academia to industry can have a lot of surprises. In an engineering school, skill sets and knowledge are relatively uniform whereas workplace knowledge is often heterogeneous and unpredictable. As such, a lot of my work is helping data scientists and next MLOps practitioners in sane development practices so ML Use Cases can be put into production. Consider the classic MLOps situation: You are working with data scientists and data analysts who deliver insights into your data and answer questions. However, that doesn’t mean the models they use are production-ready: there are a variety of challenges that MLOps work on so that ML solutions can be put into production. You might receive a non-runnable notebook with little to no documentation, or you might even receive passable code with untraceable dependencies. How does one make sure works can be productionnalized? It all stems from building common knowledge and establishing good practices to prevent problems.