Snapchat Machine Learning Engineer Interview Experience — Two Rounds on Video Retrieval and CLIP, Then Silence

I just finished two rounds of interviews this week, applying machine learning to design a short-video recommendation system.

The first round covered how to design a query retrieval system, assuming only 20% of the short videos have text descriptions. They specifically asked how CLIP works, what loss functions are used in contrastive learning, what the drawbacks of embedding-based methods are, whether there are other methods, and how to address popularity bias.

The second round covered ML fundamentals, then a project from my resume, followed by some rapid-fire questions.

Overall, I probably didn't answer the first round very well, especially the question about the drawbacks of embedding retrieval and how to address them. I don't think I answered the disadvantages of embedding retrieval very well either. It's been several days with no news, so I've already moved on.