1. Googleyness & Leadership interview
  2. Coding, algorithms, data structures, CS fundamental (Could be general or AI coding)
  3. ML /NLP design interview (One could be a bit more general design, but I will direct the interviews to focus more on ML/NLP)


  • handwriteK-MeanshandwriteKNN(eg2NN)andpushthelossfunctionoflogisticregression(nocodingrequired)Thepreparationmethodisalsoveryviolentwhichistomemorizeit


  • Designing a machine learning (ML) system to solve a crossword puzzle is a challenging yet intriguing task. Let’s break down the process into steps for clarity:
  • Certainly! Let’s structure your template for an ML design interview:

Title: Summary of ML Design Interview Strategies


  • Sharing a set of ML design answer templates based on successful experiences in major manufacturers’ interviews.
  • Emphasizing mutual benefit and knowledge sharing.

Key Strategies:

  1. Clarify Core Issues:
    • Understand the core of the problem.
    • Identify the type of ML task (classification, regression, relevance/matching/ranking).
    • Determine the model type needed.
  2. Visual Representation:
    • Use a whiteboard to draw a workflow diagram.
    • Illustrate the logical relationship of the solution’s components.
  3. Discuss the Model:
    • Engage in an interactive discussion with the interviewer.
    • Highlight key components in your model frame.
    • Analyze pros and cons of different models.
    • Visualize model structures (e.g., layers in a DNN, logistic regression optimization).
    • Show depth in understanding through detailed explanations.
  4. Evaluation:
    • Discuss evaluation metrics: ROC/AUC curve, domain-specific metrics (e.g., CTR for advertisements), confusion matrix.
    • Extend discussion to precision, recall, accuracy, etc.

Bonus Points:

  • Demonstrate proficiency in parameter estimation and compare optimization methods (e.g., MSE, log-likelihood, GD, SGD, ADAM).
  • Lead the conversation and proactively cover all logical parts.
  • Use diagrams to guide the discussion.
  • Ask for feedback on any missed points.


  • Reiterate the importance of mutual benefit and knowledge sharing.
  • Emphasize prompt and efficient communication during the interview.

  • Designing a system like Google Maps with a focus on route planning involves both ML and system design components. Given the complexity and the breadth of this topic, it’s important to structure your answer efficiently. Here’s how you can approach this in an interview setting:
    • GNN, ETA
  • Certainly, let’s space this out and delve into each of the four scenarios:
  1. Yelp Restaurant Recommendations Involving Geolocation Information
    • This involves using geolocation data to recommend restaurants.
    • The recommendation algorithm can consider the user’s current location, preferred cuisines, and past dining history.
    • It might also factor in ratings and reviews from other users in similar locations.
  2. Facebook Newsfeed Recommendations Involving Different Users’ Previous Networks
    • The recommendation system here utilizes the user’s social network and interaction history.
    • It involves analyzing the types of posts the user interacts with, the friends and pages they follow, and their network’s activity.
    • The goal is to personalize the newsfeed to show the most relevant and engaging content.
  3. Instagram Story Recommendations (Each Story is Unique and Time-Sensitive)
    • Instagram Story recommendations need to account for the ephemeral and time-sensitive nature of Stories.
    • The algorithm could prioritize Stories from close connections and those with content similar to what the user frequently engages with.
    • Time decay factors are crucial here, as Stories are only available for a short period.
  4. Spotify Music Recommendation (How to Make an Embedding of Music)
    • Music recommendation in Spotify involves creating embeddings of music tracks.
    • This could be based on the audio features of the music (like genre, tempo, rhythm), user listening history, and collaborative filtering.
    • The challenge is to capture the essence of a music track in a multi-dimensional vector that can be used to find similar tracks or predict user preferences.
  • Notification from Recruiter: The recruiter notified that there was a down-level to L4. Also, there was a request for assistance from the recruiting team.
  • First Round - Behavior: The focus was on questions related to Inclusion.
  • Second Round - ML Design: This round included questions about ML and NLP for half an hour. The initial question was about Expectation-Maximization (EM), and there was also a query about implementing WordPiece. Most questions were basic and usual. The second half included a question on Text Classification.
  • Third Round - Coding: The round started with a question about designing a 64-bit Readable Timestamp in the database and using the extra bits. This was followed by a coding test related to Index, which was not seen on LeetCode but was of medium to hard difficulty. It was completed successfully.
  • Fourth Round - ML/NLP Knowledge/Design: This was a design problem involving two thousand categories. The specifics of the question are not clearly remembered, but it was assessed as medium to hard difficulty. The task was completed in twenty minutes, ending the interview early.
  • Overall Reflection: The interview did not go very smoothly. Despite reading about interview experiences and preparing for a routine approach, some aspects were unexpected.

  • This round included questions about ML and NLP for half an hour. The initial question was about Expectation-Maximization (EM), and there was also a query about implementing WordPiece. Most questions were basic and usual. The second half included a question on Text Classification.
  • Absolutely, let’s dive into more detailed explanations of each topic:
  1. What is ROC-AUC?
    • ROC-AUC Explained: The Receiver Operating Characteristic (ROC) curve is a graphical plot that illustrates the diagnostic ability of a binary classifier as its discrimination threshold is varied. The Area Under the Curve (AUC) represents the measure of the ability of a classifier to distinguish between classes. The higher the AUC, the better the model is at predicting 0s as 0s and 1s as 1s.
    • Usage: To calculate ROC-AUC, you plot the True Positive Rate (TPR) against the False Positive Rate (FPR) at various threshold settings. The TPR is on the y-axis, and the FPR is on the x-axis.
  2. How to Deal with Imbalanced Datasets
    • Techniques:
      • Oversampling the Minority Class: Increasing the number of instances from the underrepresented class in the dataset. For instance, using the Synthetic Minority Over-sampling Technique (SMOTE).
      • Undersampling the Majority Class: Reducing the number of instances from the overrepresented class.
      • Algorithmic Adjustments: Utilizing algorithms that are inherently equipped to handle imbalances, like using tree-based algorithms.
    • Evaluation Metrics: Focus on metrics like Precision-Recall AUC instead of accuracy, as accuracy can be misleading in imbalanced datasets.
  3. Questions on Feature Engineering
    • Techniques:
      • Feature Selection: Selecting the most relevant features for use in model construction. Techniques include filter methods, wrapper methods, and embedded methods.
      • Feature Extraction: Transforming the data into a lower-dimensional space (e.g., PCA - Principal Component Analysis).
      • Feature Construction: Creating new features based on the existing ones (e.g., creating polynomial features from linear ones).
    • Implementation: Tools like scikit-learn in Python offer various utilities for feature selection and extraction.
  4. How Many Models Can Solve This Problem? (Pros & Cons)
    • Different Models: Linear Regression, Logistic Regression, Decision Trees, Random Forests, Support Vector Machines, Neural Networks, etc.
    • Pros & Cons: Each model has its strengths and weaknesses. For instance, linear models are simple and interpretable but may not capture complex relationships. Neural networks are powerful for capturing nonlinearities but require more data and are less interpretable.
  5. How to Do Online/Offline Evaluation
    • Offline Evaluation: Involves testing a model on a pre-collected dataset. It’s used for initial model testing and tuning. Methods include train-test split, cross-validation, etc.
    • Online Evaluation: Involves evaluating a model in a live environment. Common methods are A/B testing where two versions of a model are run simultaneously on different user groups to compare performance.
    • Metrics: Depending on the problem, you might use accuracy, precision, recall, F1 score, etc., for offline evaluation. For online evaluation, metrics like click-through rates, conversion rates, or user engagement metrics are common.
  • By combining these detailed explanations with practical examples and implementations (especially in popular ML frameworks like scikit-learn, TensorFlow, or PyTorch), you can form a comprehensive understanding of each aspect of machine learning.
  • How would you design Google Docs? How would you design Google Home (voice assistant)? How would you design Amazon’s books preview? How would you design a social network? How would you design a task scheduling system? How would you design a ticketing platform? How would you design a system that counts the number of clicks on YouTube videos? How would you design a webpage that can show the status of 10M+ users including: name, photo, badge and points? How would you design a function that schedules jobs on a rack of machines knowing that each job requires a certain amount of CPU & RAM, and each machine has different amounts of CPU & RAM? Multiple jobs can be scheduled on the same machine as long as it can support it.
  • General

How would you build, train, and deploy a system that detects if multimedia and/or ad content violates terms or contains offensive materials? Design autocomplete and/or spell check on a mobile device Design autocomplete and/or automatic responses for email Design the YouTube recommendation system Follow-up questions

How would you optimize prediction throughput for a RNN based model? What loss function will you optimize and why? What data will you collect to train your model and why? How will you avoid bias and feedback loops? How will you handle a corrupt model or an incorrect training batch?