Today, companies dedicate significant resources to implementing ML models, only to see a myriad of unexpected performance degradation issues arise when they are deployed.
To overcome these challenges, data organizations are increasingly turning to ML engineers to help bridge the gap between the data scientists that build the models and the teams that operationalize them.
While Lyft is a household name, the average rider may not know that the company has facilitated more than a billion rides worldwide.
The backbone of the Lyft app is a state-of-the-art ML infrastructure that is used to deliver solutions ranging from mapping and fraud detection to pricing optimization and ETA estimates.
We recently spoke with Alex Zamoshchin, engineering manager at Lyft, to understand how the ML engineer is evolving to meet the needs of the company’s critical ML initiatives.
“At the highest level, Lyft relies on ML engineers to help get models from research into the real world while ensuring they achieve business objectives,” says Zamoshchin. “They’re involved in everything from framing ML problems within the business context and converting models into working pipelines to analyzing experimental and observational data to ensure model quality and performance once deployed.”
Zamoshchin, who previously worked as an ML engineer at Palantir Technologies, now manages Lyft’s Driver Dispatch team, responsible for the models that match more than one million riders and drivers across the Lyft Marketplace each day.
“The underlying system that decides how to dispatch each ride is tremendously complex and relies on a diverse team,” adds Zamoshchin. “The process of identifying a need, translating it into a mathematical framing and implementing, observing and ensuring the performance of the resulting model often means you have contributors to this process that don’t speak the same technical language.”
ML engineers, according to Zamoshchin, play an essential role in bridging the gap between the teams that research and build and those responsible for putting them into production.
In a hypothetical world without ML engineers, Zamoshchin describes a scenario where the data science team has an idea, builds a model, and then relies on the production team to implement it. Because data scientists are removed from production environments and their limitations, their models may require features that don’t exist or wouldn’t be worth the effort to bring them online. As a result, the model goes back to the drawing board or doesn’t perform as planned if it’s deployed.
In reality, says Zamoshchin, ML engineers help resolve potential issues with models before and after they are implemented, avoiding the undesirable situation where models are impossible to implement or fail in the wild.
While ML engineers can come from various backgrounds, they are the common ground amongst all the contributors across an ML team.
“We have ML engineers who are data scientists trained in statistics and others from a computer science background. We have team members who are experts in deep learning and others specializing in maintaining and debugging models. They come from all across the spectrum,” closes Zamoshchin. “At the end of the day; however, ML engineers possess the full set of skills required to build ML workflows and infrastructure and smoothly move projects from inception to production.”