went through figma's ML engineer loop a few months ago. didn't take the offer (went somewhere else) but it was a solid process. sharing the breakdown because there's almost nothing out there on what the figma ML interview actually looks like.
context: the ML team at figma is relatively small and product-embedded. they're not building foundational models. they're applying ML to product problems: auto-layout suggestions, font matching, search ranking, smart animate, AI-assisted design features. knowing this shapes what they actually test.
what the loop looked like for me: recruiter screen ML-specific phone screen with an ML eng (60 min) virtual onsite: 4 rounds
phone screen: a mix of ML fundamentals and one coding problem. they asked about precision vs. recall tradeoffs in the context of a recommendation system, not just the textbook definition. then a medium LC problem. pretty standard, nothing exotic.
onsite round 1 - ML design: this is the most important round. i was asked to design a system that recommends color palettes to users based on their current design. we got into feature engineering (what signals from the canvas would you use?), model selection, how to handle cold start, evaluation metrics. they want to see product thinking layered on top of the ML. pure modeling depth is not enough.
onsite round 2 - coding: two problems. one algorithmic (arrays/hashing), one more applied. the applied problem involved processing a data stream and making a running prediction. not a typical LC problem. more like something you'd actually write at work.
onsite round 3 - ML depth: deeper technical questions. this is where they probe whether you actually know your stuff. we talked about transformer architectures at a conceptual level, when to use contrastive learning, and i got a question about handling distribution shift in a deployed model. they didn't expect me to go extremely deep on every topic but they noticed quickly when i was BSing vs. when i had real intuition.
onsite round 4 - behavioral: typical behavioral questions, but with a figma twist: they asked how i'd work with the design team to define what 'good' looks like for an ML feature. answers that assumed pure engineering ownership did not land well based on my read of the interviewer reactions.
overall: solid interview, harder than i expected on the product-ML intersection, lighter on pure research depth than i expected. if you're coming from a pure research background you'll need to show you can translate.