finished ibm's DS loop last month. targeting an AI/ML product analytics role under the watsonx umbrella. the loop had four rounds, spread over two weeks.
sql round: this is the most well-defined round. they give you a schema (i got something resembling a customer event table with session logs and product usage), and you write four or five queries of increasing complexity. window functions, CTEs, a self-join on the harder end. nothing crazy but it wasn't 'select count(*)' either.
tip: read every query before you write it. i nearly missed a filtering condition on Q3 because i started typing too fast.
stats and probability: average treatment effect, A/B test power calculation, 'how would you detect if a metric is behaving anomalously in a time series'. i got a p-value interpretation question which sounds trivial but they actually wanted me to explain what it doesn't mean (significance != effect size, etc.). if you've been lazy on stats fundamentals like i have sometimes, brush up.
case / product analytics: 'IBM has an enterprise client whose model accuracy has dropped 8% over the last quarter. walk me through how you'd diagnose this.' starts broad, gets more specific as you ask clarifying questions. they're testing whether you'd check for data drift, distribution shift in input features, training pipeline changes, upstream schema changes. i missed one obvious branch (recent data labeling process change) and they prompted me, which is fine.
ml / modeling round: not as deep as you'd expect. more 'explain why you'd use gradient boosting vs. logistic regression for X' than 'derive the backprop equations.' feature engineering discussion, how you'd handle class imbalance, model deployment and monitoring.
bottom line: SQL is table stakes, stats needs to be solid, the case round is actually the differentiator. a lot of DS candidates who are strong on modeling fall apart on the structured diagnosis piece.