Blind has helped me a ton throughout my interview journey! It can be hard to find information for machine learning positions (ie machine learning engineer, applied scientist, data scientist), so I want to share my experience interviewing for ML positions. All roles were primarily for MLE and I am currently an applied scientist at Microsoft. This post will discuss the companies I interviewed with, the MLE interview process, and my preparation. Happy to answer any questions! I interviewed for around 2 months and because I interviewed with many companies, I had 1-2 phone screens every day and 1-3 onsites every week at some points. The mental load of managing all this was very stressful. I did this while still working, albeit work was more chill so I had time to interview. In the end I had received several offers and still had a lot of onsites left, but I was satisfied with my offers and cancelled the rest of my onsites. I want to reiterate that I never thought I could even get a single offer let alone more than one! Luck plays a big part in the interview process. I've had interviews which I thought went well only to get rejected, and I've had interviews which I thought went poorly only to get an offer. Location: SF Bay Area YOE: 3 + masters Current TC: $260k (after done vesting $180k) *** Companies: *** Offers: Scale AI L4 (mid-level), Base $220k, 85k options at $4 strike price TC $376k Bytedance 2-1 (senior), Base $230k, RSU $215k/4 yrs, Target Bonus $56k (25%), Sign-on $40k TC $340k Twitter L5 (mid-level), Base $180k, RSU $380k/4 yrs, Target Bonus $27k (15%), Sign-on $35k TC $310k Square L5 (senior), Base $180k, RSU $500k/4 yr, Sign-on: $10k TC $305k Discord L2 (mid-level) Not sharing breakdown here for privacy, feel free to DM if curious Approximate TC $300k Amazon L4 Applied Scientist (entry-level), Base $200k, RSU $230k 5/15/40/40, Sign-on Yr 1: $115k, Sign-on Yr 2: 90k TC $330k Onsites: Instacart - Rejected Phone Screens: Pinterest - Passed Airbnb - Passed Uber - Passed Doordash - Passed Twitch - Passed Reddit - Passed Tinder - Passed Snap - Rejected Redfin - Rejected LinkedIn - Rejected Netflix - Rejected Online Assessment: OpenAI - Rejected Ultimately chose Discord for the best career growth and domain alignment. *** MLE Interview Process: *** Phone Screens: For most companies this consisted of a mix of the following. Usually interviewers will pick 2 out of the below 3. Most often I got asked to talk through a past project + 1 coding question. - Talk through a past ML project: The interviewer will ask about the data, labelling, modeling, and evaluation metrics of an ML project you did in the past - Coding: Mostly standard LC problems, some interviewers like to ask for pseudocode on implementing gradient descent. Usually only 1 question will be asked. - ML Theory: Questions about ML concepts like explaining model tradeoffs, when to use which eval metric, how to deal with unbalanced datasets, model interpretability, how to improve a model, when to productionize a model. This was mostly about traditional ML not much about deep learning. Onsites: Most companies will have the following rounds: - Behavioral - Coding: Same as above, often there were 2 coding questions asked - ML Theory: Same as above, but there may be more deep dives into your specialty (ie NLP, CV) - ML Case Study / ML System Design: You are given a vague product problem like improving engagement on a feed, improve customer churn, return items from search engine query, etc. You need to translate it into an ML problem like text classification, recommender systems, churn detection, ranker, etc. You are expected to clarify requirements, talk through potential datasets, possible features and how you would featurize them, model tradeoffs, offline and online evaluation metrics, and model productionization. Usually you are given some direction on what data is available but you can make up any dataset you want to use. Speaking about model productionization and scaling is an important aspect especially for senior levels. Uncommon rounds but exist: - ML Coding: Given a dataset, implement a simple model training/evaluation using sklearn, Tensorflow, PyTorch, etc. I has this round with 2 companies. It can be difficult to not run into data or model bugs which are difficult to debug. - SDE System Design: A standard scalable software system interview, ie design TinyURL, design a notification system. I had this round with 2 companies. *** Preparation Process: *** - Used LC Premium for top tagged and Blind 75 for 2 months. Probably did around 200 LC questions, practiced 1-4 questions a day. (Coding interviews were hardest for me) - Reviewed ML theory from books/blogs (ML theory interviews were easiest for me). - Topics to study include: popular traditional ML models (implementation, loss functions, tradeoffs), evaluation metrics (interpretation, tradeoffs), model interpretability, sampling techniques, techniques for unbalanced datasets, gradient descent/optimization, if NLP (text pre-processing, tf-idf, transformers). - As I did more interviews I learned what knowledge gaps I had, then studied those topics. - Prepared ML system design with Grokking ML System Design, blog posts, Youtube videos, and mocks with friends. You should probably study on examples like recommendation systems, feed ranking, classification problems (binary and multi), regression problems (I was not asked this by any company though). #tech #google #amazon #meta #apple #netflix #microsoft #linkedin #interview #ama #machinelearning #machinelearningengineer #offers #uber
Wow very informative, thanks for sharing! What about probability and stats questions, such as “what’s the average number of times you can roll a 6-sided die without seeing the same number twice?” I hate those things ☹️
I was sometimes asked questions about different types of distributions and Maximum Likelihood Estimator during the ML theory interviews. I wasn't asked pure probability/stats questions actually. I think if you apply for Data Scientist positions which lean towards the analytics side you would be asked that.
I see. Yes, I have a book I’ve been using that’s for DS interview prep and that’s where I’ve been seeing this stuff. For the Maximum Likelihood, are we talking general knowledge questions or like derive the MLE for this weird thing? I have my first on-site this week and still trying to figure out what is the most relevant material to study.
For the Twitter offer, is L5 identical to SWE II?
Yes it is MLE 2
Thanks for the info. Do you think 1 month of prep is enough for ML Design? I have theoretical ML knowledge but never done ML Design before.
Yes 1 month should be ok. And yeah use grokking ML. I also like reading Eugene Yan’s blog for some more ML design for recsys
What is grokking ml? Link?
This is great information! We’re you approached by recruiters, apply directly, or both?
A mix of both. I would say 50/50 split
Much appreciated. Good to know it’s not a waste of time applying directly - I’ve read some folks are not having much success in doing so.
What did the ask you in openai? And which position you applied for? And did you try for DeepMind?
Research Engineer. It was a code signal assessment very heavy on math programming with numpy. I did not apply for DeepMind.
Thanks for sharing this and congratulations on your offers. Which companies asked you the traditional SDE system design rounds
Twitter and Discord. I believe Reddit would too since their ML hiring process doesn’t seem that mature yet. My phone screen quizzed me on plain SDE material.
Thanks for your response. Your experience suggests that MLE/AS candidates can better invest their time prepping for ML design interviews instead of system design. Would you agree?
Congrats!
hi OP, I have a 3rd round hiring manager interview with ByteDance (TikTok) coming soon. would you mind to tell me how was it for your case? I did 1st round - coding & ML knowledge 2nd round - coding & ML sys. design so far. I'm curious if there would be another coding, or I should focus more on ml sys. design part
It was behavioral with 1 easy coding question.
thank you for replying op. was it DP question? also, is it more like technical + behavioral? (technical detail of previous projects?)
Hey OP, congrats on the offers and thanks for the write up. I’ve got my on-site with Square for a DS position coming up, any insight into what kind of business question/ML model they had you build?
Fraud detection classifier
Amazon AS n MLE in different companies interviews are the same?
Yes Amazon and Microsoft AS interviews are the same as other companies' MLE interviews. On the job though Amazon and Microsoft AS focus more on model development less on building normal software systems whereas other companies are more likely to have more software engineering work for MLEs.
Uber also has AS I guess. Any idea why G n Meta don’t have AS roles?