Meta Machine Learning Engineer Interview Guide

This guide uses the internal Meta MLE interview docs compiled by their own machine learning engineers as a bible (which we have thanks to our coaches). The round names and focus areas for each in this guide are verbatim and also feature advice from Prepfully’s coach network, including Staff and Senior MLEs. 

The interview loop officially goes through these stages:

1. Recruiter screen
2. Technical Screen
3. Onsite interviews:
   2 x Coding
   AI-Enabled coding (replaces 1 traditional coding round if applicable)
   ML system design
   Behavioral
4. Hiring Committee and Team Match

If you’re targeting multiple MLE offers, read through our Google, Amazon and Microsoft MLE interview guides to better understand their expectations and interview process. 

This is a 30 minute call on which you'll talk to your recruiter about your

• background
• experience
• long-term goals, 
• why Meta 

and other relevant details so they can lock in a level that sets the bar for how you’ll be evaluated for the rest of the loop.

They'll also ask you about your areas of expertise and a few questions on how you've addressed technical challenges in some of your previous projects.

You can expect questions like:

• Have you worked with large datasets? If so, how did you handle data preprocessing and ensure data quality?
• Have you implemented any machine learning algorithms from scratch? If so, can you provide an example and explain your thought process?

There are 2 things to do on this call:

• Confirm the coding language preference well in advance for the technical rounds and the details of the loop you’re about to face (coding formats, how many design rounds, etc.)
• Establish a good rapport with your recruiter by being upfront about your background and what you want out of your career. Meta recruiters are pretty much famous for advocating for candidates they like later in the process, so they're good people to have on your side. 

We also recommend that you try and deflect any comp questions because your final level will be decided at the end of the interview loop by the hiring committee and comp will be decided accordingly.

This is a live coding interview conducted via a collaborative editor like CoderPad with a Meta engineer. Meta says that it’s formatted to see how you can quickly come up with the most efficient solution to a problem in your head first, and then quickly code through it without any logic flaws. It lasts 45 minutes, and the time split is roughly:

• 5 min: introductions
• 35 mins: coding
• 5 mins: your questions

The code execution is switched off so every solution is checked by reading it. 

For the screen, the topics you face are sorted into 

1. Critical tier: 

• Structures: Arrays/lists, BSTs, dicts/hash maps, stacks/queues, graphs
• Algorithms: Binary search, BFS/DFS, quicksort, mergesort
• Concepts: Big O, recursion

1. Nice-to-have tier

• Structures: Tries, PQs/heaps, sets, RBTs
• Algorithms: DP, MSTs/min-cut, radix sort, heap sort, randomized quicksort

Officially, this screen is judged against 8 considerations:

• Listening carefully and comprehending the question
• Asking the right clarifying questions 
• Heeding to hints 
• Speed in solving problems 
• Surveying multiple approaches to solve before committing to one 
• Looking for new ideas/methods while solving 
• Being inventive/flexible if new data/constraints are introduced 
• Ability to move onto more complex problem solving  

Candidates interviewing at E6 describe a slightly different screen where there's a coding block of about 25 minutes and there's a dedicated 15 minute behavioral segment folded in.

For the screen, expect problems that are of medium difficulty and prioritize optimal time/space complexity like:

• Find the LCA of 2 nodes in a standard binary tree
• Find the k closest points to the origin (0,0) on a 2D plane, given an array of coordinates.
• Compute the exact frequency of a target integer in a sorted array in O(logN) time

Time can be a little tight in this round, but you should take about 2-5 minutes to ask the interviewer about corner cases. Our coaches also say that it's better to have a non-optimal but working solution than random fragments of an optimal but an incomplete one.

The entire onsite loop will have about four to six rounds of 45 minute interviews that’ll cover

• Coding
• ML system design
• Behavioral

Typically, a candidate faces 2 coding rounds, and since October 2025, one of them is replaced by the AI-Enabled coding interview, which we'll discuss shortly.

This round tests your DSA and code execution skills in 2 LeetCode style medium-hard difficulty problems. The round runs in CoderPad where the interviewer won't compile your code. You can write in any language you're comfortable with, though our coaches say it’s just easier to go with scripting ones. 

The interviewers are usually thinking about how your skills and experience might help Meta, and also how you tackle problems you're not necessarily familiar with. They'll be scoring 4 focus areas:

• Problem Solving: the quality of your reasoning behind a solution, developing and comparing more than one and choosing the appropriate data structures and algos and speaking to space and time complexity on the way to optimizing.
• Coding: Assesses if you can convert solutions into executable code and if the code is organized and captures the right logical structure.
• Verification: this is a check on if you're considering a reasonable number of test cases or coming up with a good argument for why your code is correct. This aspect is pretty important because you’re expected to manually trace edge cases and you don’t have the luxury of a compiler or autocomplete. The official Meta guide that we mentioned earlier states very clearly that catching a bug is good news (and even uses an exclamation point). 
• Communication: This is a fairly direct assessment of if you're asking the right questions about reqs and keeping the interviewer in the loop about your approach. 

Expect questions like: 

• Decide whether a string can become a palindrome after deleting at most one character
• Simplify a Unix-style file path. (A reported variation starts from a current directory, applies a relative path, and asks for the navigation from the current location to the target)
• Rearrange an array of integers into the next lexicographically greater permutation

For practice, our coaches recommend fewer example problems of many problem types over drilling into a single type. There's a fair bit of overlap for coding problems between software engineers and machine learning engineers so anything tagged as SWE is good practice too.

Also, watch Meta’s Hacking the Coding interview. 

This is a 60 minute AI assistant coding round where you'll work in a three panel CoderPad layout. The left panel is a file explorer, the middle is the code editor, and the one on the right contains problem instructions and the AI assistant chat window.

You'll get to choose the model you want.

The AI has full context of the repo’s files, but it can't modify code directly inline - it can only reply in chat. You’ll have to modify the code yourself.

The whole interview centers on a single cohesive project or an applied scenario like building game mechanics or simulating a spreadsheet engine. It typically goes through 3 phases

1. Exploring and fixing issues in an unfamiliar codebase
2. Implementing a new functionality like a new feature or extending a core function
3. Optimizing/refining the code for runtime behavior or interpreting mock data, etc. 

Meta describes this round as “more representative of the developer environment that our future employees will work in” so they're also very clear about what they're assessing you for.

• Critical thinking v/s blind trust: The lightweight LLMs frequently hallucinate or output some buggy code. So if you dump a giant prompt into the chat and blindly paste it without inspecting it, it's not the best look on your part
• Verification: Essentially, you should be able to isolate test failures. Run a couple of test cases individually rather than all at once to show that the AI's logic works before you commit to it.
• Communication: Treat the AI like a junior engineer and talk to your interviewer about why you're framing your prompt a certain way and where you think the AI might fall short.

Some tips for this round from our experts:

• Resist the urge to dive into a prompt immediately. Take a few minutes to read the data models and look over the existing files.
• Going by ground reality, many candidates don't finish phase 3, but still pass. High quality validation and strong logic in phases 1 and 2 matters way more so don’t rush through the process only to end up with a sloppy finish.

In this round, you're given an open-ended product prompt to architect a large-scale production-ready machine learning service from end to end using a virtual whiteboard (like Excalidraw). 

The number of ML design rounds depends on your level. E4 and E5 candidates usually get 1, and E6 gets 2. More and more E5s have reported getting 2 of these.

The five focus areas carry the weight of the round:

• Problem Navigation: A read on if you can visualize and organize the entire problem and solution space, and connect the business needs to ML decisions
• Training Data: How you meticulously collect training data and how you look at the constraints and risks within a proposed method.
• Feature Engineering: Checks if your features are relevant and how you'd identify the most important features for that specific task.
• Modeling: If you're able to justify your choice of model and explain how it's trained, you're also expected to anticipate the risks that come with it and the ways to mitigate them.
• Evaluation & Deployment: This asks for consistent evaluation and deployment techniques on your end and justification for the metrics you want to track.

Expect questions like:

• Design the recommendation and ranking system for Instagram Reels.
• Design the ranking system for Facebook News Feed.
• Design an ads ranking or click-prediction system.
• Design a system that detects harmful content.

Generally speaking, as intuitive as they seem, these are the tips we would like to give to you: 

• Brush up on basic ML theory and algorithm details. Be comfortable with concepts like overfitting and regularization.
• Practice converting intuitive ideas to concrete features. For example, "number of likes" can be a good feature suggestion, but a better feature might also involve normalization, smoothing, and bucketing.
• Having a good tool set of several different algorithms and understanding the tradeoffs is helpful. For example, be able to explain the advantages of logistic regression compared to SVM.
• Design from the ground up. Think about how you'd design a system that Meta (or another large tech company) already has. It's a good exercise to think through the complicated, high-scale systems that you already use every day.

There are 3 things to note here:

• Some candidates reported a traditional system design round along with their ML system design. If you want to brush up on those details, you should visit our system design guide.
• Going by postmortems, this round carries a lot of weight in the leveling decision. A candidate interviewing as an E5 MLE with strong feedback everywhere except ML design was offered 3 paths later: an E4 ML, an E5 SWE, or a follow-up ML design interview.
• Finally, since 2026, this round has started focusing on sub-domains specific to what the company’s looking most for. For instance, a significant number of candidates have reported their ML Design actually focused entirely on GenAI ML Design / LLM Design. Those interviewing in Reality Labs focused on Computer Vision (CV) focused ML Design. A few (decreasing) number of candidates reported traditional ML Design focused on recommenders, ranking etc. Please definitely check with your recruiter if your interviewer will be focusing on a subdomain because it makes a meaningful difference in your choice at architecture etc; and therefore your preparation as well. The good news is: Prepfully ML coaches who’ve worked at Meta cover pretty much all domains between them; but always make sure you provide the domain when booking a coach so that they can decline if they don’t support a specific one.

In this session, the interviewer will want to learn about your background and what you're passionate about in tech. Expect 4-5 questions with follow ups for each. They'll try to assess if you thrive in Meta’s fast-paced and highly unstructured environment, and that'll be assessed by gauging 5 signals.

• Resolving conflict: how you view/manage disagreements with colleagues and if you can empathize with the people whose points differ from yours radically.
• Growing Continuously: actively seeking out and valuing chances to grow and learn professionally 
• Embracing ambiguity: showing resourcefulness and a bias for action even when the path/project is not defined clearly
• Driving Results: Pushing yourself and others to deliver against your shared goals
• Communicating Effectively: Being able to keep all stakeholders in the loop by tailoring your communication based on the work and audience

Expect questions like:

• Have you ever hurt a working relationship because of a conflict? Did it recover?
• How did you react when a project’s priority shifted unexpectedly?
• Have you ever taken the lead on a project and then realized you had incomplete data?
• How do you keep your xfn partners (DSs, PMs, etc) accountable for all their deliverables when you’re working on a project?

When you sit down to build a story bank, we recommend that you take the classic STAR framework and build it out into STAR+ (Situation, Task, Action, Result + the X-Factor: Learning) to structure your stories. The X-Factor at the end - Learning - is your definitive chance to demonstrate intellectual humility, and helps you reflect on a personal takeaway on how that situation changed you for the better. 

It’s good to prep these before your loop starts because E5+ candidates reported behavioral questioning in their tech screen too. 

At the end of each round, interviewers get 24 hours to submit their notes and pick between Strong Hire, Hire, Leaning No Hire, or No Hire for you. 

The Hiring Committee, consisting of senior MLEs, then reviews that feedback to make the final call and decide your level. 

If the committee signs off on a “yes”, you’ll then move into the Team Matching phase where you'll do 30-minute informal chats with HMs until you zero in on a fit.

E3: $187K

E4: $295K

E5: $453K

E6: $659K

View the latest compensation for Meta MLEs on Levels.fyi