Operationalizing Machine Learning — Unified Approach

Last week we shared some thoughts about the pain-points of operationalizing machine learning across industries, surveyed and summarized by a group of Berkley researchers led by Shreya Shankar.

The consensus found in this research motivates the next generation of solutions in the area of MLOps.

Today the data stacks are more fragmented than unified, as exemplified in our previous post regarding data science tooling. More and more data and MLOps tools are being built that aims to fill a niche gap somewhere along the data science workflow. On the other hand, there’s a call for unification of tools so as to simplify the DS tech stack as much as possible.

I want to contrast two schools of thoughts in the unification of tools. One was the previously mentioned “decoupling” engineering from AI. I borrowed the idea from Harvard Professor Stratos Idreos who’s devoting his research to disrupting the MLOps industry and providing a unified solution (I’ll write more about it in a future post!).

The second was the Nbdev approach of enabling a better Jupyter Notebook. Instead of removing the engineering component, this approach aims to make it easy for data scientists to incorporate the engineering pipelines within the environment that they are familiar with, synchronizing the best of both…