The recruiter sent a HackerRank OA first. It had two standard LeetCode questions and some multiple-choice questions on ML fundamentals.

Then there was a 45-minute video screen with a LeetCode medium and some quick questions on gradient descent and overfitting.

The onsite was four rounds and honestly pretty intense:

• One pure coding round where I think you should expect optimized solutions, as brute force won't help.
• One ML theory round going deep into the math behind my past projects, embeddings, attention mechanisms, and ranking metrics like NDCG.
• The real filter is the ML system design round, where I had to design a recommendation engine from scratch that revolves around real-time inference, candidate generation versus ranking, and serving at a massive scale.
• The last round is behavioral, mostly about standard past experience.

For preparation, I read through Chip Huyen's ML systems book, which gave a solid baseline. ML design is so open-ended, though, so I really wanted to do a dry run. I managed to find a TikTok MLE on Prepfully for a mock after searching a lot on other sites. It was totally worth it, and he showed me how to actually structure the recommendation pipeline on a whiteboard without getting stuck on irrelevant details.