Square · Primly Community

Square machine learning engineer interview: the full loop, including the parts nobody warns you about

ml_mike · 4 replies

Went through the Square MLE interview loop this spring for a role on their risk/fraud ML team. Going to be blunt: it's a harder loop than people expect for a company that isn't typically on the 'FAANG gauntlet' tier. Here's what happened.

Six rounds total: Recruiter screen ML fundamentals (60 min) Coding round (LC-style, 60 min) ML system design (60 min) Behavioral (45 min) Hiring manager depth conversation

ML fundamentals: The interviewer went deep. Questions ranged from gradient descent and regularization to imbalanced classification problems, which is extremely relevant to fraud detection. They asked how I'd handle extreme class imbalance (fraud rates are like 0.1%), what metrics I'd use beyond accuracy, and how I'd explain model decisions to non-technical stakeholders. The stakeholder explainability angle caught me a bit off guard. I gave an okay answer but I could have been more specific about SHAP values and partial dependence plots.

Coding: Two medium-ish problems. One was algorithmic (sliding window pattern), one felt like a data cleaning problem that you'd actually see in a real job. Reasonable. Nothing tricky.

ML system design: This is the round I'd tell anyone to prep hardest. My prompt was: design a real-time transaction fraud detection system. I had to cover feature engineering (including some real-time feature stores, which I'd worked with before), model training pipeline, inference latency requirements, monitoring for model drift. The interviewer pushed hard on the tradeoff between a fast online model vs. a more powerful batch model. Time constraints for payment networks matter here: you have milliseconds at the card swipe layer.

The fraud detection domain knowledge mattered. Not that you had to have worked at Square before, but vague ML hand-waving wouldn't land. Be specific about feature engineering choices, latency vs. precision tradeoffs, and how you'd monitor for concept drift in a financial context.

Offer was competitive: my total comp was around $260k for L5 equivalent in SF, base plus RSUs on a standard 4-year vest. That's roughly in line with what I'd expect for the level at a non-FAANG fintech in 2026.

4 replies

pivot_pat

The imbalanced classification question is a gimme for fraud teams but I've watched senior candidates fumble it. Oversampling vs. undersampling vs. threshold adjustment vs. cost-sensitive learning: you should have crisp opinions on when each applies, not just list them.

mobile_mara

Good point on the stakeholder explainability angle. That's increasingly a question in regulated domains. Being able to explain a model decision to a compliance team or a customer flagged as fraudulent is a real operational need, not just a fluffy 'communication skills' question.

sec_sasha

Latency at the card network layer is genuinely brutal. Depending on the integration point you might have under 100ms to make a call. System design that ignores that is going to read as unrealistic to anyone who's worked in payments.

ml_mike

Yeah sec_sasha is right. I got pushed on this specifically. Had to articulate why I'd use a lightweight online model for real-time decisions and kick the heavier ensemble to an async layer for post-authorization review. The interviewer seemed satisfied with that framing.