Brex · Primly Community

Brex data scientist interview (SQL + case + stats), what the loop looked like

ds_dmitri · 6 replies

Went through a Brex DS interview loop targeting a mid-level data scientist role in Q1 2026. Four rounds total. Going to break this down because the mix was different from what I expected.

Round 1: SQL Two questions, roughly 45 minutes. The first was a window functions problem on a transaction dataset, something like "find customers whose spend increased more than 20% month over month for three consecutive months." Tricky to write cleanly but not algorithmically hard. The second was a funnel analysis query. Know your CTEs, window functions, and how to calculate conversion rates in SQL.

Round 2: Stats + Probability This was more rigorous than I expected for a startup. They asked a Bayesian inference question (not trick-question style, but genuinely applied: "your fraud model flags a transaction, the model has X precision and Y recall, how do you think about the posterior probability of actual fraud?"). Also a question about A/B testing assumptions, specifically whether the standard z-test is appropriate when your metric is revenue per user at a company with heavy right-tailed distributions.

Round 3: Case / Product Sense They gave me a hypothetical scenario: Brex card spend dropped 15% last month. Walk through how you'd diagnose it. This is classic but they pushed hard on how I'd distinguish internal product issues from macroeconomic signals and from a data collection problem. The data quality angle is where a lot of DS candidates thin out.

Round 4: Behavioral Standard, focused on cross-functional collaboration and how you've influenced product decisions with data.

Salary for mid-level DS in SF market was in the $170-210k TC range based on what I've seen, though I can't share my specific number.

The SQL and stats rounds were the real filters. If you're weak on either, prep those first.

6 replies

analyst_ana

The Bayesian fraud question is so much more interesting than the typical "explain p-values" I usually get. Did they expect you to use formal notation or was it more of a conceptual discussion?

ds_dmitri

Conceptual, but precise. I wrote out the formula on the virtual whiteboard but the real conversation was about reasoning. They wanted to see that I understood base rates matter (fraud is rare), not just that I could recite Bayes' theorem.

growth_gabe

The 15% spend drop diagnostic is a good interview question honestly. Do you know if they want a structured framework answer or more of a free-form investigation?

ds_dmitri

Structured, but they don't want you reciting MECE blindly. I used a loose funnel (is it a data problem? is it a product problem? is it external?) and they kept probing each branch. The data quality check first instinct scored well.

finance_faye

Did they test domain knowledge about card spend specifically? Like would someone without fintech background be at a disadvantage?

ds_dmitri

Some. Knowing what merchant category codes are and how interchange works helps. Not required but you get points for it. Read the Brex help docs on their card product before your case round.