Ramp · Primly Community

Ramp machine learning engineer interview: what they actually test (not what you'd expect)

ml_mike · 4 replies

Finished a Ramp MLE loop in early 2026. Sharing because there wasn't much specific info out there when I was prepping and I had to rely on pretty general startup MLE interview guides.

First thing to know: Ramp's ML scope is narrower than FAANG but arguably more immediately impactful. Their main ML surface areas are fraud detection, spend categorization, and some forecasting work. So the interview isn't going to test you on ranking systems or recommendation engines. It's fraud-and-classification shaped.

What they tested:

Coding: One round was straightforward ML-adjacent coding. Write a function to compute precision and recall from scratch, no sklearn. Then extend it to handle multi-class. Not hard, but the point was: do you actually understand what the metric means, or do you just call .fit(). I talked through class imbalance handling in fraud contexts while I coded and they seemed to like that.

ML system design: This was the meat of it. "Design a system to detect fraudulent expense submissions." Very relevant to what they do. I covered feature engineering (merchant history, velocity, deviation from user baseline), labeling strategy (rule-based bootstrap vs human review loop), model choice (gradient boosting for interpretability in a compliance-sensitive context), threshold calibration, monitoring for distribution shift. They pushed hard on: how do you handle feedback delay? Fraud labels arrive weeks late. That's a real problem and they wanted a real answer.

Math/stats: Not a whiteboard-math interrogation but some quick questions on confusion matrices, why AUC alone can mislead on imbalanced data, what you'd use instead.

Behavioral: One dedicated behavioral round. Standard structure but Ramp-flavored: tell me about a model you shipped that didn't work as expected. What did you learn. How did you communicate to non-technical stakeholders.

Verdict: The interview is well-designed for what they're actually building. If you've worked on fraud, risk, or anomaly detection anywhere in fintech, you'll feel at home. If you've only done NLP or recsys, spend a week reading about imbalanced classification and feedback loops before you go in.

4 replies

marketer_mei

The feedback delay problem is a great signal for how mature their ML thinking is. Most early-stage fintech teams either ignore it or solve it naively. Did they have a specific approach they wanted or were they testing if you'd surface the issue at all?

ml_mike

Both, kind of. They wanted to see you surface it without prompting, and then they'd push you on specifics. I talked about delayed labeling via pseudo-label proxies (immediate rule-based labels for obvious cases, real labels flowing in later via a label pipeline) and they asked about the risks of that. They weren't looking for one correct answer; they were seeing if you'd reason through tradeoffs.

infra_ines

Any sense of the stack? Python obviously but are they on Spark, Airflow, Ray, something else?

ml_mike

Didn't get deep into infra specifics during the interview itself but from what I picked up from the team chat, Python throughout, some dbt and Snowflake on the data side. Not a heavy Spark shop from what I could tell.