Atlassian · Primly Community

Atlassian machine learning engineer interview: real breakdown of what they assess

ml_mike · 5 replies

i went through the atlassian ML engineer interview loop earlier this year. there's almost nothing written about this track so here's what i found.

atlassian has ML roles spread across a few product areas, primarily recommendation systems (for jira/confluence), search relevance, and some automation intelligence work. the role i applied for was on the search team.

the process recruiter call. take-home coding assessment (this is unusual, it was a 2-3 hour ML-focused problem, not leetcode). then virtual onsite with 4 rounds.

take-home they gave a dataset and asked me to train a classifier, evaluate it, and write up what i'd do differently with more time. pretty standard data science vibes but they read the writeup carefully, the onsite debrief referenced it.

onsite rounds ML depth: basically a conversation about the take-home, then moved into how i'd approach a recommendation problem from scratch. talked about candidate generation, ranking, feature engineering, offline vs online eval. they were not impressed by buzzwords. they asked follow-ups that assumed you actually understand the math. Coding: general SWE round, LC medium. no ML-specific code. System design: design a search relevance system for a document retrieval problem (very relevant to their actual product). we talked about BM25 vs dense retrieval, hybrid approaches, latency tradeoffs, how to A/B test a ranking change without wrecking the user experience. Behavioral: atlassian values again. same weight as all the other positions.

what they actually care about honestly it felt less ML-theory-heavy than big tech and more about engineering judgment. can you make good decisions about model selection given real constraints. one interviewer specifically asked "what would you cut first if you had to ship in half the time."

took about 7 weeks total. offer did come through. happy to answer questions.

5 replies

ds_dmitri

the take-home is interesting. did they give you a time limit or just a rough guideline? and what format did you submit, a jupyter notebook or something else?

ml_mike

they said "spend around 2-3 hours" but there's no enforcement. i submitted a notebook with a short written summary at the top. the quality of the writeup matters as much as the model results, in my experience they're more interested in your reasoning than your AUC.

infra_ines

the search relevance system design is pretty on-brand for atlassian. confluence and jira search have been notoriously rough for years. makes sense they'd probe into that.

remote_swe_42

what level did you end up at? curious how atlassian levels MLE vs big tech. and was it remote?

ml_mike

came in at what maps roughly to a senior IC, atlassian uses P5/P6 style leveling internally i think. remote US. total comp on the offer was around 260k TC which is below google/meta MLE rates but the culture and velocity were the draw for me.