Round 1 (Practical engineering + debugging + discussion):
I had to write a simplified environment interface for an agentic RL/evaluation scenario and debug concurrent execution code, with the focus on state isolation and exception handling. I did not finish.
Round 2 (ML evaluation coding):
Live Python coding: implement a classic model-calibration metric, Expected Calibration Error, over sharded data and get the tests to pass. The main focus was code cleanliness and handling edge cases.
The questions themselves were quite close to real post-training and evaluation work, but my personal experience was just so-so. The atmosphere in the first round was especially pressured. The interviewer kept following up and even questioning whether I was really familiar with agent harnesses. It was not a comfortable conversation.
Discussion
Loading comments…