Machine Learning Engineer · Full Journey · Multiple Types

Meta — Machine Learning Engineer ✅ Passed

Level: Senior-Level

Round: Full Journey · Type: Multiple Types · Difficulty: 6/10 · Duration: 240 min · Interviewer: Unfriendly

Topics: Probability, String Manipulation, Algorithms, AI-Enabled Coding, Behavioral Questions, Machine Learning System Design, Recommendation Systems, Feature Engineering, Exploration Strategies, Cold Start, Position Bias, Training/Serving Skew, Feedback Loops

Location: San Francisco Bay Area

Interview date: 2025-12-23

Summary

Interview Rounds Overview

• Round 1: Phone Screen
• Round 2: Coding
• Round 3: AI
• Round 4: Behavioral
• Round 5: ML Architecture & Design

Details

I interviewed for a Machine Learning Engineer role. I received an E5 offer and proceeded to team matching.

Phone Screen:

I had a 45-minute phone screen with two LeetCode medium questions:

• The first question was probability-related, similar to LeetCode 528, Random Pick with Weight. My algorithmic approach was correct, using cumulative sum and binary search, but I had a bug in my implementation.
• The second question was string manipulation, similar to LeetCode 1249, Minimum Remove to Make Valid Parentheses. I needed a hint from the interviewer to solve it, and my implementation had minor bugs.

The interviewer passed me despite my perceived poor performance. It was likely due to my correct approach, clear communication, and willingness to accept hints.

Virtual Onsite:

The virtual onsite consisted of four rounds over three days.

• Round 1: Coding (45 minutes)

  Two questions:

  • LeetCode 680, Valid Palindrome II
  • LeetCode 215, Kth Largest Element in an Array

  I solved both questions quickly, as I had practiced them extensively. This round was straightforward.

• Round 2: AI-Enabled Coding (60 minutes)

  This round used Meta's new AI-assisted format.

  • Question: Maze Solver

    Four progressive problems:

    • Q1 (no AI assistance): Took about 25 minutes; I got stuck.
    • Q2-Q4 (with AI assistance): I completed all of them.

  I had specifically prepared for maze problems beforehand, which helped me complete the later parts smoothly. Most candidates only finished 2-3 parts, while completing all four was considered a top performance.

• Round 3: Behavioral (45 minutes)

  I used the STAR method to answer several behavioral questions. I rated my performance as 7/10. The content of my stories was good, but the delivery could have been more concise.

• Round 4: ML Architecture & Design (45 minutes)

  Question: Short video recommendation system (similar to TikTok/Reels)

  My solution:

  • Retrieval: Used a Two-Tower model for candidate generation.
  • Ranking: DCN-V2 / MMoE multi-objective (P(click), P(comment), P(share), E(watch_percentage)).
  • Re-ranking: Business logic + diversity considerations.
  • Feature engineering: Video features, user features, cross features, GraphSAGE social features, context features.
  • Exploration: Contextual bandits (epsilon-greedy + Thompson sampling, 2% exploration traffic).
  • Cold start handling.

  Follow-up questions covered: position bias (using inverse propensity weighting), training/serving skew, and feedback loop issues.

  Overall, I felt my performance was strong E5, borderline E6.

LeetCode similar: LeetCode 528, LeetCode 1249, LeetCode 680, LeetCode 215