Studying real-world ML applications helps you think like a senior engineer.
Search
Ranking Relevance
BERT, RankBrain
NLP, Query Intent
Photos
Searchable Images
CNNs, Vision Transformers
Image Embedding
YouTube
Recommendations
Two-Tower Models
Retrieval & Ranking
Gmail
Spam/Smart Reply
RNNs, LSTMs, Transformers
Sequence Learning
YouTube uses two main networks:
Candidate Generation: Filters millions of videos to hundreds using user history.
Ranking: Assigns a precise score to each candidate to produce a final list.
Metric: Not just CTR (Click-Through Rate), but Watch Time (engagement quality).
Uses powerful visual embeddings.
Task: Identify people/objects without explicit tags.
Tech: Models like CLIP or contrastive learning to align images and text in a shared space.
1. "If you were designing a 'Related Articles' feature for Google News, what would be your primary metric?"
While CTR is easy to measure, User Satisfaction or Retention (do they come back tomorrow?) is the long-term goal. Metrics like "Time Spent on Article" or "Share Rate" are better proxies for quality than just clicks.
2. "How would you handle 'Fairness' in a recommendation system?"
This is a frequent Google-style question. Discuss Unbiased Data (removing sensitive attributes), Equity in exposure for minority creators, and monitoring for echo chambers.
3. "Why does Google Search use multiple models for one query?"
Complexity. One model handles spelling, another handles synonyms, another handles semantic intent (BERT), and finally, a ranking model combines all signals. This is called a Modular Architecture.
When answering case study questions:
Scope the problem (Users, scale, goal).
Handle Data (Where do labels come from?).
Draft Architecture (Input $\rightarrow$ Model $\rightarrow$ Output).
Discuss Edge Cases (Cold start, malicious users).
Last updated