First round was Recruiter round where i was asked questions about my background and experience specific to their role and expectations. I was asked to write a simple code as well.

The company scheduled 4 hours of tests with no time dedicated to answer questions, or try to get you interested (typically, this is 30% of the interview process). The interviewers would run out the time and leave on the hour, and emailed questions were not answered to any informative degree. Information on the upcoming interviews was very conflicted - "the next interview is Leetcode"; in fact it was ML/pandas. The second round of interviews was particularly bad. I was told it would be Pytorch debugging. A person out of grad school who was not qualified led the interview. The model got 0.69 train accuracy, and I was supposed to get val accuracy up. I pointed this out as underfitting went looking for the bug. She kept pulling me off saying there is no bug. "What kind of layers can you add? What about some transforms? What are the data labels telling you?" These were obvious hints for dropout, augmentation, and class balance, but they do nothing to address poor underfitting which I was pointing out is the problem. After the interview, I did find that the model was calculating loss as log_softmax(log_softmax(softmax)). Removing the inner layers gets 0.95 train accuracy, at which point it makes sense to look at generalization. So, the interviewer failed their own question...