did the McKinsey MLE loop earlier this year for a role in their QuantumBlack practice. want to share what the interview actually looked like vs what i expected.
i expected heavy leetcode and some standard ML theory. what i got was much more applied and communication-heavy. take that as good news or bad news depending on what you're strong at.
the rounds
five total. recruiter call, technical screen, three onsite sessions (ML technical, system design, PEI + case).
ML technical round
model building: they gave me a business scenario (churn prediction for a telecom client) and walked through the whole ML workflow as a conversation. feature engineering decisions. how do you handle class imbalance. train/val/test split strategy. how do you evaluate the model beyond accuracy (they explicitly asked this, precision/recall trade-off mattered here). they asked about XGBoost vs logistic regression for this use case and wanted real reasoning, not just 'XGBoost is better.'
one question on model monitoring: 'you've deployed this churn model, six months later performance has degraded. how do you diagnose and fix it.' i walked through distribution shift, feature drift, label drift, pipeline issues. they seemed satisfied.
no implementation coding. no LeetCode. the one coding question was 'write a function that computes precision and recall from a list of predictions and labels.' like, from scratch without importing sklearn. straightforward.
system design (ML systems)
how do you build an ML training pipeline that runs weekly retraining on fresh data. cover the feature store, model versioning, A/B testing infrastructure, rollback. i drew this out in a shared doc. they pushed hard on 'what happens when training data has corrupted records' and 'how do you version models alongside the training data they came from.'
the case + PEI
yep, even for MLE. the case was a lite version: 'a client's recommendation model is hurting conversion vs their old rule-based system, why might that be.' this is actually a great ML question disguised as a case. i had a real answer (distribution shift, feedback loop, position bias in training data). it landed well.
PEI: had two stories ready, used them both. they drilled down hard on impact.
bottom line
if you're a strong applied ML person who can communicate, McKinsey is genuinely interesting. if you want to optimize LeetCode for two months and coast into a job, look elsewhere.