Citadel · Primly Community

Citadel machine learning engineer interview: what they actually tested and where i struggled

ml_mike · 5 replies

interviewed for an MLE role at citadel quantitative strategies earlier this year. this was for a senior-level position, so take that lens into account. sharing what i actually remember, not a sanitized version.

the recruiter screen: short, maybe 20 minutes. they asked about my ML background, which domains i'd worked in, and whether i was comfortable with both research and production. mentioned that the role was heavily production-focused, not pure research.

technical screen 1: ML systems + coding this was the most interesting round. they gave me a problem where i had to design and implement a simple feature pipeline for a time-series prediction problem. not a full training loop, just the feature engineering piece. i wrote it in python. they asked why i made certain choices around windowing, normalization, and handling missing data. then they asked me to walk through how i'd take this to production: serving latency, retraining cadence, monitoring for distribution shift.

if you haven't thought carefully about monitoring and drift detection, this is where you'll feel the gap.

technical screen 2: applied ML concepts no coding. more like a conversation. they asked about specific model families: why you'd pick gradient boosting over a neural net for a given problem, how to handle class imbalance, what you'd do when your model degrades after 6 weeks in prod. one question i didn't expect: how do you explain a model decision to a trader who doesn't care about ML?

onsite: four rounds. one was a full system design for an ML platform feature (online learning, near real-time inference). one was a coding round closer to a data structures / algo problem. one was a deep dive on a project i'd shipped. and one was behavioral.

the coding round felt a bit out of place honestly. it was more leetcode-adjacent than the rest of the loop, which was very applied. if you're coming from pure research, prep your CS fundamentals more than you might expect.

comp for the level i was at: base around $265k, bonus component that's significant but variable. total package landed well above most big tech ML roles but the bonus is genuinely not guaranteed.

citadel MLE interviews are hard but they test real things. not much leetcode theater, mostly whether you can reason about production ML under pressure.

5 replies

ds_dmitri

the 'explain it to a trader' question is such a good signal. that shows up in finance ML interviews a lot and a lot of ML people bomb it because they go deep on SHAP values when the person just wants to know if they should trust the signal or not.

ml_mike

exactly. i ended up framing it as: here's the one or two input features that drive most of the prediction in this scenario, here's the direction, here's the historical accuracy. kept it to what a non-technical stakeholder actually needs. seemed to land well.

infra_ines

the drift detection question is where people get caught. everyone knows to say 'monitor it' but if you can't talk about specific signals (covariate shift vs label shift, what you'd alert on) it's obvious pretty fast.

newgrad_neil

is this role open to new grads or is citadel only realistic for 3+ YOE in ML?

ml_mike

they do hire new grads into their quantitative researcher track but that's a different loop, more math-heavy. the MLE track i went through really does expect production experience. i wouldn't target it out of undergrad.