I applied through a recruiter for the Amazon personalization team.

There was an HR screening and initial tech screening before the loop. I got rejected after the tech screening.

The tech screening was at a high level. I answered most of the questions correctly, but I don't know what went wrong.

What is the most impactful project you've worked on recently, and how did it work end-to-end?

What systematic process do you use to evaluate and iteratively improve prompts for complex LLM workflows?

Beyond offline test sets, did you perform any online or live evaluations before deployment?

What are the structural and functional differences between encoder-only and decoder-only architectures?

How does a decoder model process and learn from an input sequence token by token?

How is cross-entropy loss mathematically formulated?

How does multi-head attention operate in parallel, and how are the outputs from individual heads combined before passing to the feed-forward network?