LinkedIn Staff Machine Learning Engineer Interview Experience — Phone Screen Plus a Five-Round Onsite Full of Probability Questions

The phone screen was 1hr, there was a lot of content, but it was all questions that had already shown up in other interview reports.

Coding - biased coin
Input: A method getRandom01Biased() that generates a random integer in [0, 1], where 0 is generated with probability p and 1 is generated with probability (1-p)
Output: A method getRandom06Uniform() that generates a random integer in [0, 6] with uniform probability

BQs were the standard questions.

There was also a mini ML system design question: build a LinkedIn Learning recommendation — how do you recommend courses based on the job a user applied to. Note that you're supposed to mention how to extract features from the job post — I didn't go in that direction at first, and the interviewer specifically hinted at it. There was basically no time to deep dive, so we quickly went through the high level and the feature engineering.

Onsite

My questions were basically the same as another interview report I'd seen on the forum, though the follow-ups weren't quite the same — too bad it happened to be the same day I was interviewing.

Coding1: Given a probability distribution, simulate an M-sided die that outputs a random index — hand-write binary search.
Follow-up 1: What if the probabilities don't add up to 1? We discussed two approaches — one is to scale, the other is to do rejection sampling at the random step. That second approach kind of dug a hole for myself: the interviewer followed up asking what the average number of random draws would be. I wrote out the formula but couldn't remember the value. The interviewer then just told me to google it lol — the answer is 2.
Follow-up 2: What if the probability list is very large, with the probability mass concentrated on just a few values? We discussed an engineering solution for this, with pros and cons.
There was still time left, so the interviewer also asked me how to sample uniformly from the area of a circle. I honestly didn't know this one, so I just went with my gut and gave a rough direction. The interviewer said it was correct, and that was it. Felt like something he came up with on the spot.

Coding2: Same as the second question from the report mentioned above, just needed to write it faster. I spent a long time explaining the monotonic stack solution to the interviewer, which meant I almost didn't finish writing the code — I think there were some edge cases I didn't handle well, and I didn't have time to check.

ML foundation: The interviewer started from a design problem. I don't remember the exact question, but there were a lot of features mentioned. We then discussed how to do feature selection, plus a lot of ML basics and practical questions.
Does the model's output probability need calibration, and how do you calibrate it?
Feature selection — how do you do it with a neural network?
Then we got onto tokenizers.
Asked about optimizers — I'd forgotten the details of Adam.

ML design: Design LinkedIn Skills. I drew a diagram, walked through the flow, and afterward we mainly discussed how to collect data, feature engineering, and how to apply LLMs.

BQ: Standard BQ questions, but afterward there was a rec design problem where we didn't really discuss technical details — the manager kept asking me how I'd plan it, how I'd collaborate, roughly how many people I'd need, and how long it would take. Seems like they wanted to test whether I could lead a large project. I hadn't really prepared well for this beforehand, so I just answered based on gut feeling.

Overall, I think LinkedIn's MLE role is actually a mix of applied science and MLE, and they seem to have a bit of an obsession with probability questions haha. The staff-level bar felt like a 5.5–6 — you need to be able to both act as TL and manage a team.