Anthropic MLE loop. Full breakdown.
Signed an MLE offer at Anthropic last week. 4 rounds total. Recruiter screen plus 3 onsite. Going round by round.
Recruiter screen, 30 min. Standard fit + comp + timeline. She pushed once on a 4 month gap in late 2024 (post-burnout, was honest about it).
Round 1, coding, 60 min, one problem. Online stream processing. Wanted a function that maintained a moving aggregate over an unbounded stream with bounded memory. The signature design was half the round. Clean impl in 35, then he extended it to handle stream rate exceeding consumer rate. Backpressure discussion took the rest.
Round 2, ML system design, 60 min. RAG for a domain with rapidly changing facts. 15 min on reqs (p95 200ms, 1h freshness window, eval setup). Settled on hybrid retrieval with a dedicated freshness index. He pushed hard on dataset poisoning (10 min) and on the eval-drift loop. Strongest round of the day for me.
Round 3, behavioral + research depth, 60 min. Half STAR, half "walk me through a paper or project you wish you'd done differently." Had the story, got pushed on whether I'd actually run it the way I described in hindsight. Second half is harder than people prep for, you have to be honest about the flaws in something you built.
How I used cluely: rounds 1 and 2, not the behavioral. On round 1 it had the streaming pattern surfaced before I finished restating the prompt, which kept me out of a windowing dead end. On round 2 it surfaced the freshness vs consistency tradeoff while I was still writing reqs, which is the framing they actually graded on. Off for behavioral because the prep work for that round is your own stories and the overlay doesn't have those.
Friction: first 5 min of every round it's reading context, I carried the opens. On round 2 the interviewer pulled up a whiteboard tool I'd never seen and the screen share context got confused. Fell back to talking it through.
Offer came 6 days after the final. Happy to go deeper on any round.
