Did the Pinterest ML engineer loop back in April for a role on their recommendation systems team. Sharing what I found, because there's not much out there that's current.
Pinterest's ML org is serious. They've published a ton of papers on real-time personalization, two-tower retrieval, and multi-task ranking. If you're interviewing for a recsys or ranking role, read at least two of their papers before you go in. Not to quote them, but because the interviewers will assume familiarity and the design questions are framed around that problem space.
The loop (5 rounds total) Recruiter / hiring manager intro ML fundamentals coding (Python) ML system design Applied ML / case study Behavioral
ML coding round This wasn't Leetcode. They gave me a dataset and asked me to implement a logistic regression from scratch (no sklearn), then discuss regularization choices. They also asked about precision/recall tradeoffs in a ranking context, which is very Pinterest. Know your metrics: NDCG, MRR, hit rate. If you've only ever chased accuracy, you'll struggle here.
ML system design I was asked to design a "related Pins" retrieval and ranking system. Two-tower encoder model, approximate nearest-neighbor indexing, re-ranking with engagement signals, feature store, online/offline pipelines. They went deep on latency: how do you serve retrieval results under 100ms at Pinterest's scale? Candidates who jump straight to "use a transformer" without talking about compute cost get knocked for it.
Applied ML / case study Given a business scenario: engagement on home feed is declining, you have a set of signals, what do you do? This is part product thinking, part experimentation, part ML. They wanted to see structured thinking, not just model architecture.
Behavioral Heavy on cross-functional influence. Pinterest ML engineers work closely with product and infra. They want to know you can communicate tradeoffs to non-ML stakeholders.
I've interviewed at a lot of recsys shops and Pinterest's technical bar is real. They're not going through the motions.