just wrapped a brex ml engineer loop and wanted to write this up while it's fresh. brex isn't an ml-first company, which means their ml interviews are shaped differently than what you'd expect from a pure ml shop like spotify or airbnb.
what the loop looked like: recruiter screen (30 min, standard) ml fundamentals screen with a senior ml engineer (60 min) coding round: data manipulation + model evaluation script in python (60 min) system design: ml-flavored (60 min) behavioral + cross-functional (45 min)
the ml fundamentals screen: not a quiz. they asked me to walk through a past project end to end: how i framed the problem, chose features, validated offline vs online. they pushed on things like class imbalance in fraud detection and how i'd think about threshold selection. brex has a payments/fraud context, so skewing toward that domain in your examples helps.
coding round: no leetcode. they gave me a csv of transaction data and asked me to compute features, identify anomalies, and write a basic evaluation loop. pandas + sklearn. the point seemed to be: can you do applied ml work, not theoretical ml math.
ml system design: they asked me to design a real-time transaction risk scoring system. latency constraints, feature stores, model serving, feedback loops. i leaned into the fintech-specific complexity (card-not-present fraud, velocity checks) and that landed well. they want engineers who understand the product domain, not just model architectures.
behavioral: normal for fintech. tell me about a time you disagreed with a business stakeholder about a model's readiness. how do you handle a model in prod that's degrading.
overall take: brex ml interviews feel more like applied engineering than pure research. if your background is heavy on transformers and papers, prepare to contextualize that to practical product impact. if you're a solid ml engineer who ships things, you'll feel comfortable here.