Mercury · Primly Community

Mercury machine learning engineer interview, what the loop looks like and whether it's worth doing

ml_mike · 4 replies

just got out of mercury's ML engineering loop, rejecting an offer from them. writing this up because ML roles at fintechs are a different animal from big tech and i want to share what i learned.

first: mercury doesn't have a huge ML org. this is a small focused team building models that touch fraud detection, credit risk signals, and product recommendations inside the dashboard. you will not be working on LLMs for their own sake. if you want that, look elsewhere.

recruiter screen: 30 min, pretty thorough for a first call. she actually asked about my specific model experience (tabular data vs. NLP vs. cv) and was upfront that the team uses tabular models heavily. i have a mostly NLP background, which she flagged. that honesty was appreciated even if it signaled a rough road.

technical phone screen (60 min): two parts. first half was a ML fundamentals conversation: overfitting, regularization, how you'd handle class imbalance in a fraud dataset (spoiler: that's not an accident). second half was a coding problem in python, not an algorithm puzzle but more like writing feature engineering code given a dataframe. think pandas/polars style manipulation.

virtual onsite (4 rounds): ML system design. i was given a fraud detection scenario and asked to walk through the full pipeline: feature engineering, model selection, serving architecture, monitoring in production. they cared a lot about the monitoring piece specifically. model drift in fraud detection is a real operational problem and they want to know if you've thought about it. coding round (more feature engineering + some sklearn). product/stakeholder round (how do you explain model decisions to non-technical people, specific examples). behavioral.

what they're really looking for: practical ML for a financial context. if you can talk about class imbalance, data leakage in time-series splits, model explainability (shap values came up naturally in my conversation), and monitoring in prod, you're well positioned. research pedigree matters less than applied chops.

comp offer was around 200-220k TC. i passed because the team size was too small for where i am in my career right now, not because of anything negative about the process.

4 replies

ds_dmitri

the time-series data leakage point is so often missed. did they ask about it directly or did it come up in your system design walkthrough?

ml_mike

came up in system design when i was describing the train/test split. i mentioned leakage proactively and the interviewer lit up a little. i think it was a flag they were looking for. if you're doing fraud detection ML and you don't know why temporal train/test splits matter you're going to have a bad time.

de_derek

the overlap between DE and MLE work at this kind of company is real. feature engineering pipelines in prod are just data pipelines with extra steps. makes sense both loops overlap on the practical data manipulation piece.

infra_ines

curious about the serving architecture question. did they want a full MLflow/SageMaker deep dive or more high-level 'how does this model get called in production'?