Saranyan - MLOps that works: How we built ML pipelines for deploying models for autonomous factories

Speaker: Saranyan Vigraham - Head of Engineering, Petuum
- Speaker Bio -
Saranyan loves building meaningful products. In the past, he has led diverse engineering teams of over one hundred engineers, rallying them around vision and engineering excellence. Saranyan has shipped both enterprise and consumer products during his tenure at companies like Petuum, Elementum, Meta and PayPal. With a PhD in Computer Science, he is always learning how to push the boundaries of technology to shape society in meaningful ways.
- Talk Abstract -
ML Ops is currently is as much an art as much a science. Unlike Devops, which can be codified to a set of tools and practices to ensure consistent and efficient software delivery, the trial and error nature of ML projects pose a different set of challenges. At Petuum, we have deployed many ML models, which autonomously operate factory equipment. Because the factory environments are dynamic, it required us to build flexible pipelines that allow easy iteration and deployment, sometimes testing multiple models concurrently. In this talk, I will present one such ML pipeline (powered by an enterprise platform) that has allowed us to deploy hundreds of models. I will discuss some design/architectural patterns and anti-patterns based on our real-world experiences.