Sony · Primly Community

Sony machine learning engineer interview: recsys focus, what they actually test

ml_mike · 4 replies

Sony has a few different ML teams and the interview experience varies significantly depending on which one you're talking to. i interviewed for an ML engineer role on the PlayStation recommendation team (think: game discovery, personalization) in early 2026. sharing what i saw.

the recruiter was upfront about the team's stack: they're running a mix of pytorch-based models and classical collaborative filtering, serving on GCP. the work is genuinely recsys-heavy, not LLM fine-tuning or the trendy stuff. good if you want that, less exciting if you want gen-ai work.

process: recruiter screen (30 min): background, visa status (they sponsor), rough timeline. quick. technical screen with ML engineer (60 min): first 20 min was probability and statistics. conditional probability, bayesian reasoning, bias/variance tradeoff. standard but real. next 30 min was a recsys problem: given sparse interaction data and cold-start users, how would you approach recommendations. i talked through collaborative filtering vs content-based, how i'd handle the cold-start problem, and evaluation metrics (NDCG, precision@k). they pushed back on my evaluation choices in a healthy way. last 10 min was one coding problem in python, data structure level, easy. onsite (4 rounds, virtual): ML system design: design a personalized game recommendation system for PS5 from scratch, including data pipeline, feature engineering, model serving latency constraints. this was 45 minutes and they went deep on the serving side, not just modeling. coding: two problems, medium difficulty, both on array/hash-map patterns. not ML-specific, just general DSA. ML depth: explain a model you've built in production. walk through training, evaluation, deployment, monitoring. i used an item-to-item similarity model from a previous job. they drilled on how i detected model drift and what my retraining cadence was. behavioral: two stories. one about influencing without authority, one about a technical decision you made that you later regretted.

overall: solid loop, real technical depth on the ML side. the serving latency question caught me a bit flat-footed. if you're interviewing for their recsys team, know your inference optimization basics, batching, caching, latency budgets.

4 replies

ae_andre

the NDCG vs MAP distinction matters so much in recsys interviews and almost nobody prepares for it because it's not in the standard ml interview prep guides. what comp range did they share for the ML eng role?

marketer_mei

interesting that they're not heavy on the LLM side. for a gaming/entertainment company i'd have expected more interest in gen-ai features. did the team seem excited about the work or more in maintenance mode?

ml_mike

honestly somewhere in between. the personalization problems they're working on are real and interesting, but this isn't a research team and they were clear about that. if you want to publish papers this is not the place. if you want to ship a model that millions of people use to find their next game, that's what they're offering.

content_cole

the serving latency question is real. i've seen ml candidates who are great at modeling get tripped up because they've never had to think about p99 latency or what happens when your feature store is 200ms round trip. that's infrastructure knowledge and it's increasingly expected.