Roblox · Primly Community

Roblox machine learning engineer interview, what they actually care about vs what you'd expect

ml_mike · 4 replies

Did the Roblox MLE loop a few months back. Got the offer, didn't take it (team fit wasn't right for me), but the process was interesting enough to write up.

Roblox ML is mostly recsys and content moderation. Their main ML surface is the discovery feed (games/experiences), ad ranking, and safety/trust systems. If you come in expecting to talk about LLMs you might be surprised. It's very much classical recommendation systems and classification.

The rounds:

Coding: One LC-style round, medium difficulty. Nothing ML-specific in the coding round, just general DS+A. I got a graph problem (shortest path variant). Pretty standard.

ML design: This is the real interview. I was asked to design a recommendation system for the Roblox homepage. It needed to handle cold start (new users with no history), balance engagement vs. creator diversity, and work at scale (they have hundreds of millions of MAU). They care a lot about how you think about the objective function, what proxy metrics you'd optimize for in stage 1, and what you'd watch as your leading indicator of success. Make sure you can talk about embedding models, two-tower architectures, and approximate nearest neighbor search. Also: online vs. offline evaluation, A/B testing design.

ML breadth: One round that was more conversational. They asked about my experience with different model types, feature engineering for sparse data, and a case study from my past work. 'Tell me about a model you shipped and what happened after you shipped it' came up. They want to hear about monitoring, drift, failure modes.

Behavioral: Pretty normal. One specific question was 'tell me about a time you had to push back on a product decision because the ML wasn't ready.' That's a real Roblox thing, apparently. Their product teams want fast iteration and ML needs more runway, so that tension is real.

Comp for my offer (senior MLE level, San Mateo): Base was $195k, RSUs over 4 years. Total comp year 1 was in the $280-300k range with the equity tranche. Reasonable for the Bay Area but behind Google/Meta for the same level.

If you're coming from ads-recsys or content safety ML, this is a natural fit. If your background is purely CV or LLM finetuning, expect to do some translation work in the design round.

4 replies

hardware_hugo

The cold-start problem is always the first thing I ask when I do MLE design rounds too. How did you handle it in your answer? Did they want something model-based or were they fine with a simpler heuristic for cold start?

ml_mike

I suggested a tiered approach: heuristic fallback for zero-history users (trending + genre-based from onboarding quiz), then transition to a lightweight collaborative filter once you have even a few interactions. They liked that I called out that you can't evaluate cold-start with the same offline metric as warm users and that you need a separate holdout. The key was admitting the heuristic isn't 'ML' and being clear about when you'd transition.

consultant_cam

Appreciate the comp data. The $280-300k year 1 number is consistent with what I've seen from other Roblox MLE offers in early 2026. Worth noting that's including the 1-year cliff equity tranche, so year 2 numbers can look different depending on stock performance.

ops_omar

Did they go deep on the infrastructure side at all? Like how models get served, latency budgets, that kind of thing? Or was it purely modeling?