just got out of the jane street ML engineer interview loop. going to be direct: this is not a standard ML interview and if you prep like it is, you'll be underprepared in some areas and overprepared in others.
what's different:
no ML system design in the typical sense. they didn't ask me to design a recommendation engine or an ad ranking system. instead the design question was closer to: here's a modeling problem with financial time series data, walk me through your approach. very applied, very domain-specific. you need to understand things like signal decay, non-stationarity, and why a model that backtests well might not perform live.
coding is harder than most ML loops. two of my rounds were pure algorithms/data structures, basically senior SWE level. no "implement gradient descent" type questions. they want ML engineers who can code, not ML engineers who primarily use notebooks.
math is real. one round had probability and stats questions. not trick brainteasers but genuine applied probability. conditional probability, distributions, the kind of stuff that matters when you're thinking about uncertainty in a live trading environment. i was asked a question about expected value that had a non-obvious setup.
what didn't come up: transformer architecture deep dives, LLM fine-tuning, MLOps/model deployment pipelines (surprising given the role title).
rounds: recruiter screen take-home (data analysis in python, python was fine, no notebooks required) two live coding rounds applied ML/stats round system design (see above)
honest take: if you're coming from an LLM/NLP/applied AI product company and your strength is prompt engineering or fine-tuning, this will be a rough fit. if you came up doing signal processing, time series, or anything quant-adjacent, you'll feel more at home.
did not take the offer (different reasons, not the process), but it was a rigorous and fair evaluation.