Visa · Primly Community

Visa machine learning engineer interview: rounds, questions, and what they're actually testing

ml_mike · 6 replies

just came out the other side of a Visa MLE interview for their fraud and risk ML team. posting because the info out there is thin and what exists is a couple years old.

background: i applied for a senior MLE role in Foster City, Calif. my background is mostly production ML, some NLP, a lot of recsys. Visa's ML org sits under their Visa Analytics Platform and is pretty focused on fraud detection, credit risk scoring, and transaction anomaly detection. knowing that before you interview helps.

the loop: recruiter screen (30 min): standard background check, comp alignment, visa status (they do sponsor H1B). pretty quick. technical phone screen (60 min): this was split into two parts. first half was ML concepts: they asked about how gradient boosting handles class imbalance (i talked about scaleposweight in XGBoost and focal loss), how you'd monitor model drift in a production fraud model, and what precision-recall tradeoffs matter in fraud vs. recall-only contexts. the second half was one LC medium coding problem, nothing crazy. onsite (4 rounds virtual): ML system design: design a real-time fraud scoring system that processes 24,000+ transactions per second (Visa's actual scale). they wanted to discuss feature engineering from raw transaction data, model serving latency requirements, how you'd handle feature skew between train and serve, and A/B testing frameworks when you can't risk letting a bad model run. this was the hardest round. coding: two problems. one on graphs (felt like LC medium), one on data manipulation more like a pandas/SQL-style problem but in Python. they weren't grinding me on syntax. ML depth: behavioral+technical hybrid. walked through a project in depth, they dug into my feature selection choices, how i evaluated the model, what i'd do differently. felt like a technical postmortem. behavioral: standard leadership stuff. influencing without authority, handling disagreement with a PM, dealing with a project that hit data quality issues mid-sprint.

what they care about: real-time serving constraints are a big deal here. if you've never thought about P99 latency or served a model under tight SLAs, read up before your loop. also: class imbalance in fraud is a constant theme. have opinions on how to handle it.

timeline: 6 weeks from application to offer. offer was ~$195k base + modest equity for senior in Foster City. total comp maybe $225k depending on bonus.

6 replies

quietquit_quincy

the feature skew between train and serve point is huge and a lot of people gloss over it. glad they actually test on this. did they ask how you'd detect it in production or was it more about how to prevent it at design time?

ml_mike

both, honestly. i talked about logging feature distributions at serve time and comparing them to training distributions as part of the monitoring layer. they seemed to like that. the prevention angle (like consistent feature pipelines using the same code path for train and serve) also came up and they nodded at that.

brand_ben

data point noted. $195k base senior MLE Foster City 2026. that puts them roughly in line with other large fintech firms (not FAANG-adjacent, but not bad). the equity being modest tracks with what i've heard from others at Visa.

visa_vik

super valuable to know they sponsor H1B. do you know if they also support green card sponsorship or just H1B transfers? i'm on OPT and that's a real consideration for me.

ml_mike

my recruiter mentioned they do PERM sponsorship but i didn't dig into the details since i'm already on GC. i'd ask directly in the recruiter screen, they were pretty open about it.

ux_uma

24,000 transactions per second and they want real-time fraud scoring. the latency constraints on that are genuinely interesting. did they get into the tradeoffs between an ensemble model (slower but more accurate) vs a lighter model at the edge for first-pass filtering?