Stripe Machine Learning Engineer Interview Experience — Two Phone Screens, Account Matching and a Kaggle-Style Round

I recently interviewed for an MLE role at Stripe. I passed both phone screens and I'm waiting for the onsite. This was my first time interviewing for MLE, so it was a challenge for me too.

Both interviewers were really nice, the best I've had so far. They were very patient, gave hints, and even helped write some boilerplate code so I could test faster.

Round 1: coding

The problem: you're given a bunch of accounts, each with some metadata, and you have to find similar accounts.

The first two parts were pretty basic: loop over the list, compare the metadata of each pair of accounts, and if the fields match (name, for example) you decide they belong to the same person.

The third part was transitivity: if A is related to B and B is related to C, then A and C also count as related, even if A and C don't match directly. This is the standard indirect-association problem, so the idea is BFS or Union-Find to find connected components.

Round 2: ML Integration

People haven't said much about this round here, but I think the question bank is probably just this one problem: you get a dataset, very much like a Kaggle competition, and you have to work through it from scratch.

The focus isn't the model. Just using a Random Forest is enough. The focus is data handling: first EDA and preprocessing, checking whether the data has missing values, how to do imputing, and how to deal with outliers.

The interviewer dug into two specific things. One was checking whether the DataFrame has duplicates on one key or on a pair of keys. The other was how to one-hot encode categorical features and how to handle missing values in them.

After I built the model, there were quite a few follow-ups on model evaluation, including recall, precision and AUC, and they also asked in detail about calibration, specifically how you actually do calibration.

Some advice for anyone interviewing after me: for the Integration round, just practice Kaggle-style problems well.

I'm sharing this for the community, and good luck to me on the onsite.