This application season I applied for Roblox's PhD DS Summer Intern position. I didn't end up passing, but I felt the team's overall background was pretty solid, and the assessment was a bit different from the usual, so I wanted to share it here.
Timeline:
- 11/23/2024 — applied online, got the assessment the same day
- 12/19/2024 — submitted the assessment
- 02/12/2025 — HR reached out
- 02/27/2025 — virtual onsite
- 03/07/2025 — got the rejection letter
Assessment:
Four tasks: 1 coding skills, 2 problem-solving, 1 decision-making.
- Coding: most of the interview reports on the forum before this were about stats/ML Python code or multiple choice, but what I got was two case-style problems — they give you some background and some data, and you write Python code to do some basic data processing and analysis, like sorting, filtering for the max value, computing some averages, etc. I don't remember the exact questions, but it wasn't hard.
- Problem solving: two games, played on the Roblox platform itself. One was assembling a car to get it across obstacles, the other was building a production line to make bread or something, to maximize profit. I answered this part pretty badly.
- Decision making: 30 BQ multiple-choice questions in 30 minutes. I've attached some of them for reference.
Virtual Onsite:
45 minutes. HR said it would cover three parts: stats, product sense, and coding.
In the actual interview there was no real stats content at all. The main content was a lot like Meta's product sense case — and actually a lot of the team members had also jumped over from Meta.
The interviewer skipped introductions on both sides and went straight into asking about my resume — introduce a data project I'd done, including the details of the data used, the model I chose and why, and the final conclusions or implications.
Next was a product sense case: "How do we know players prefer local creators?" I asked two clarification questions: 1. How do we define local creators? 2. Is it reasonable to infer preferences from time spent on these games? For the first question the interviewer said they meant creators who make games in the players' native language. The second one also got a yes.
So I first defined the primary metric as time-per-session. At this point the interviewer asked a follow-up: why time-per-session instead of total time? I said time-per-session better reflects the future likelihood of monetizing that consumer, because it means the player is more immersed in the game experience and more likely to make in-game purchases or accept promotions (honestly my logic here wasn't very solid). Of course total time is the most direct metric, so I picked number of sessions as the secondary metric.
After that I asked if we could run an A/B test. The interviewer said no. So I answered that the tools we could use include matching, IV, and DID — starting from matching, doing the analysis at the player level, defining players who've played a local-creator game as the treatment group and those who haven't as control (I actually later realized you could also do this analysis at the game level, so I mentioned that at the end). Comparing the two groups directly like this has selection bias, so I'd match on player characteristics and on-platform behavior, then compare time-per-session between the two groups.
Last was a Python coding question: the platform is launching social features. Roughly, for every player id there's a list of requested-friend and accepted-friend ids with dates, and from this you extract the matched pairs of friends and the date they became friends.
Discussion
Loading comments…