Roblox Data Scientist Interview Experience — A Brutal OA With Robot-Building and Factory Games

First-time poster here. After I got the OA, I asked the career center for an extension but didn't get one, so I ended up squeezing it into one evening of Thanksgiving break, a week before finals deadlines. After I finished I felt like my English was terrible and my IQ had dropped :((( It seriously wrecked my mood for the rest of finals prep.

1. 25 minutes building a robot car. There were two rounds of scenarios. In each round you're given a ton of parts to choose from (you have 3 energy total, and each part costs somewhere between 0 and 1 energy), and you need to test-fit parts onto the left/right side, wheels, top, and front/back, then get the car through two different terrains. The first scenario looked like lava plus a broken bridge; the second was a grassy drone field plus a pinball bumper that launches you into the air. My English isn't great, so I had no idea what half the parts actually did. In hindsight, it probably would have helped to brush up on mechanical/car-building/rocket vocabulary beforehand — things like repulsor, coater, hose, various kinds of vats. The part icons don't make their function obvious either (aside from the really obvious ones like a propeller, and even then I couldn't read the English description). The text explaining what each part does is also really long, so I read through the specs painfully slowly — I probably spent half of each ~13-minute round just reading part descriptions, and it turned out to be a waste of time anyway. It would have been better to just try things directly, except there are so many parts to pick from that trying each one also eats up time, and the same part behaves differently depending on where you mount it (front/back/left-right/top). I was so confused. Overall it was really hard to build anything decent — I only ended up putting together 2+1.

2. 25 minutes on a factory profit-maximization task. By this point I was already a bit fried from the first task, so I didn't read the instructions carefully before jumping straight into the real questions. (Heads up: it looks like you can't go back and re-read the full instructions once you're past the tutorial, so don't skip ahead — read every step carefully before moving on, or you're stuck.) I'd read online beforehand that the best strategy is to figure out which product's route is most profitable, then only focus on the profitable ones and work backward adjusting the suppliers and the machines that produce the intermediate goods (a final product might go through 2-3 intermediate steps). The problem is that knowing the theory is one thing, but actually doing it is a mess — you have to type in numbers by hand at every single node (the whole graph might have dozens of nodes), and the routes for the four different products cross over each other, e.g. A and B sharing a raw material, then an intermediate step feeding into both B and C. Every step has a limitation, and raising that limitation raises the cost too, so it got extremely confusing — I'd change one number and then forget what I'd already changed earlier, or forget which node I'd left blank (you have to manually input the actual production quantity — I was completely lost). My suggestion: just get one product line working end to end first and don't touch the machine settings yet. Yes, you can also change each machine's production capacity, its raw material requirements, and its production time, but the cost goes up every time you do.

After finishing those, I remember there was a feedback survey asking whether you play video games, what kind, and how many hours a week. I shamelessly checked "3-4 hours a week" (that was probably an intermediate-tier option — realistically it's probably more).

3. 25 minutes, 23 behavioral questions — you're given a scenario and pick the best and worst option out of four. I'd just relax on this section and not overthink it; some of the "correct" answers didn't feel obvious to me, so I mostly went with gut instinct.

4. 50 minutes on CodeSignal, four questions, R or Python your choice.

4.1 Given an observed array, find the minimum total sample size for an A/B test (a two-sided z-test with equal-sized treatment and control groups), given the significance level, the size of the mean change you need to detect, and the power. The final answer needs to be rounded up — I don't normally round up in my day-to-day work, so I forgot to use math.ceil() and lost time there.

4.2 Diff-in-diff. Given three arrays — a period indicator [0/1], a treatment indicator [0/1], and an outcome [float] — calculate (a) the DiD estimate, and (b) a flag for whether there's a pretrend. The pretrend check was simple: given a threshold, you just check whether the absolute mean difference in the pre-period exceeds it.

4.3 Logistic regression. Given a training set of features and a binary outcome array, normalize the features first, fit a logistic regression, then rank the features by coefficient value and report the top 3 feature names. You can probably just call sklearn directly for this.

4.4 A very simple Bayes' rule question — should take about one line to solve.

This was my first OA ever and it felt so hard. I instantly lost all motivation to look for an industry job, TAT.