Coinbase · Primly Community

Coinbase machine learning engineer interview: what they're actually testing

ml_mike · 4 replies

Went through the Coinbase MLE loop for a mid-senior ML role. Five rounds. Let me break it down because the specific angles surprised me.

ML system design: This was the round I was most prepared for and it did not go the way I expected. I assumed they'd ask me to design a recommendation system or fraud detection (both obviously relevant for Coinbase). They asked me to design a transaction anomaly detection system. Good guess on my end actually but the nuances were very Coinbase-specific: how do you handle the fact that on-chain behavior is partially public, meaning you might have more signal than typical fintech? What about the regulatory constraints around using customer data for model training? Those compliance wrinkles were the actually interesting part.

Coding: Two rounds. First was LeetCode-style Python. I got a medium graph problem. Second was more applied. They gave me a dataset mock (just a description of columns and scale) and asked me to write pseudocode for a feature engineering pipeline for the anomaly detection task. Not about syntax, about demonstrating you can go from raw data to model-ready features and reason about what actually matters.

ML depth: One round was basically an ML knowledge interview. Questions covered: bias-variance tradeoff applied to a specific scenario they described, class imbalance handling (very relevant for fraud/anomaly), calibration of probability outputs. One question I didn't have a great answer for: how would you handle concept drift in a model that flags on-chain wallet behavior, given that legitimate behavior patterns change every few months as new protocols emerge. I gave a monitoring + retrain cadence answer. They probed further. I think I got partial credit.

Behavioral: One round focused on technical leadership. Not managing people. Making technical calls under ambiguity, influencing without authority, shipping models that aren't perfect but are good enough. Standard stuff but they framed it well.

Offer was around $210k base for a senior MLE in SF. Equity was significant. The comp felt like it tried to compete with FAANG without being FAANG, which is about right.

4 replies

ds_dmitri

The concept drift question for on-chain behavior is genuinely hard. The behavioral patterns of a DeFi power user in 2024 look nothing like 2022. Did they seem like they had a real answer themselves or were they just evaluating how you reasoned through it?

ml_mike

Felt like the latter. My impression was they wanted to see if I'd just say "retrain on new data" or actually engage with why this is hard. The regulatory piece makes it especially complicated. You can't just scrape arbitrary wallet behavior for training.

sec_sasha

The point about on-chain data being public but regulated is underrated. A lot of ML people don't think about the compliance wall you hit even when the data is technically accessible.

backend_bekah

Did they ask anything about model serving infra or was it all about the ML side? I've been doing more MLOps lately and curious if that comes up.