CI/CD for Machine Learning Systems
Machine learning systems are notoriously difficult to productionize. Lots of moving parts means lots of failure modes: are you serving the right model? Are you applying the same feature transformation logic in the service as you did during training? Did yesterday's retrained model deploy correctly? How can you make sure your model accuracy doesn't drift over time? This talk will step through these pain points - and more - and walk through how we adapted classic CI/CD techniques to solve them at Spotify
I am a hybrid engineer dabbling in Machine learning, infrastructure, backend services, UI and embedded systems. I have been in Spotify for over a year working in the data engineering team and have recently shifted my focus to ML Infra, working to reduce the time to market of an ML model and provide stricter guarantees. Prior to Spotify, I worked at Samsung R&D Institute Bangalore, working on the sensors team prototyping sensor based applications for the Galaxy line of smartphones and smart watches. I graduated from UMass Amherst with a degree in Computer Science specializing in Machine Learning.