Went through Notion's ML engineer interview in early 2026. Sharing this because I could not find anything useful when I was prepping. Most search results are about their SWE process.
First, context on what ML actually means at Notion. They've been pushing hard on AI features: the Ask Notion / AI assistant stuff, smart summaries, auto-categorization of blocks. So the ML team is not doing pure research. They want people who can build and ship, who understand LLM integration patterns, and who have opinions about when to use a fine-tuned model vs a retrieval-augmented system vs a prompt-engineered one.
Keep that in mind when you prep.
Process:
Recruiter screen (30 min) then an ML-specific phone screen with an engineer on the team. The phone screen was more of a technical conversation than a coding test. We talked through a feature I'd built involving recommendations, what the training pipeline looked like, how I'd evaluate it, and what I would change in retrospect. They weren't interested in whether I could recite ML fundamentals. They wanted to see whether I think about ML in the context of a real product.
Virtual onsite, 4 rounds:
ML system design (60 min): Design an AI-powered document summary feature. This is so clearly ripped from their actual product roadmap that it almost felt like a real spec review. I talked through chunking strategy for long docs, choice of embedding model, latency tradeoffs, how you'd handle updates when the underlying document changes. They pushed hard on the latency question. How long is acceptable for a summary? What's the user expectation? It became a product conversation as much as an ML one.
Coding (45 min): Standard Leetcode medium, nothing ML-specific. Slightly annoying if you're an ML person who hasn't been grinding algorithms, so prep for it anyway.
ML fundamentals + applied (60 min): This one went deep. Gradient descent, regularization, why dropout works, attention mechanisms. Then shifted to: how would you evaluate this summarization feature, what metrics make sense, how do you avoid eval gaming. The evaluation question is where I've seen people trip up. Know the difference between automated metrics and human eval, and have a point of view on when each matters.
Behavioral (45 min): Same as rest of the company. Bring STAR stories, especially around ambiguity and cross-functional work.
Overall: it was a harder loop than I expected for a company this size. If you're coming from a pure research background you'll want to practice translating what you've done into product terms. If you're coming from a big-tech applied ML team it'll feel more familiar.
Comp I saw was roughly in the range of $220-260k TC for L4/L5 equivalent in SF in 2026, depending on level. Smaller equity pool than FAANG but faster vest schedule.