Anthropic · Primly Community

Anthropic data engineer interview, pipelines and SQL: my full loop breakdown

de_derek · 5 replies

just finished my loop for a data engineer role at Anthropic and figured I'd write it up since there's basically nothing out there for this specific track.

Recruiter screen was pretty standard. 20 minutes, background questions, brief overview of the team (ML infra side). nothing surprising.

Then a technical phone screen, which is where things got specific. they gave me a SQL problem involving windowed aggregations across a time-series of model training runs. think: compute moving averages, identify anomalies in cost-per-token metrics across experiments. not trivial but also not trying to trick you. the context was very much "this is the data we actually work with."

Onsite was 4 rounds: Data modeling: design a schema to track LLM evaluation runs. they wanted to talk through how you'd handle versioning, experiment lineage, and eventually querying across thousands of evals. I went deep on slowly-changing dimensions and they seemed into it. Pipeline design: design a real-time + batch hybrid system for ingesting and storing token usage logs at scale. they pushed on consistency vs. throughput tradeoffs. Kafka, Flink, iceberg on S3 were all on the table. SQL deep dive: more complex than the phone screen. a few CTEs, a query optimization question where they gave me an execution plan and asked what I'd change. Cross-functional behavioral: how have you worked with ML engineers to understand data needs? how do you handle it when a researcher changes a schema mid-project? pretty standard stuff but they want real stories not generic answers.

my offer came back about 11 days after the onsite debrief. base was competitive with other AI labs I've seen (somewhere in the 185-230k range depending on level), plus significant equity. the total comp story at Anthropic is a bit different because it's pre-IPO restricted stock, not options, so you have to factor in liquidity differently.

overall: very technically grounded loop. nobody tried to stump me with trivia. the focus on AI/ML-specific data problems (eval logs, training metrics, cost tracking) felt genuine. if you're used to e-commerce or fintech data pipelines, spend time thinking about the ML infra use cases before you go in.

5 replies

ds_dmitri

the schema design for eval runs question is a really good signal about whether a place actually thinks carefully about data. did they care about foreign keys and normalization or was it more of a "talk me through your thinking" vibe?

de_derek

more the latter. they didn't want to debate normal forms. they wanted to see if you'd proactively think about things like: how do I query "all evals that used model version >= 2.1" six months from now. the temporal querying angle was the whole point.

infra_ines

iceberg on S3 for the batch side makes a lot of sense there. did they mention any specific orchestration tooling, or was it more tool-agnostic?

analyst_ana

how long was the full loop from recruiter screen to offer? trying to figure out if I should hold off other offers.

de_derek

about 5 weeks total. 2 weeks to get the phone screen scheduled, 1 week between phone screen and onsite, then 11 days for the offer. not the fastest I've seen but they were communicative throughout.