Perplexity AI·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

Senior

May 2026

Summary

Interviewed for an ML engineer role at Perplexity AI and got hit with a system design question about building a search system that keeps up with shifting user interests over time. Pretty fitting for what they actually do, which made it feel less abstract than most design questions.

Questions Asked (1)

How would you design a search system that continuously adapts to users' evolving interests?

System DesignTechnical Trade-offsProduct Strategy

Author's notes

This one has a lot of surface area and I probably tried to cover too much.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the system's scope and requirements, then propose a modular architecture that separates real-time personalization from offline model updates. Emphasize a feedback loop where user interactions continuously inform ranking and retrieval models, and discuss trade-offs between freshness, relevance, and computational cost.

Pro tip: Highlight the importance of measuring long-term user satisfaction (e.g., via retention or session depth) rather than just click-through rate, and mention how you'd guard against feedback loops that reinforce narrow interests.

1. Clarify Requirements and Constraints

Ask about scale, latency, data availability, and business goals (e.g., freshness vs. diversity). Establish what 'evolving interests' means in this context.

2. Design Data Pipeline and Feature Store

Outline how to collect and process user interactions (clicks, dwell time, queries) in real-time and batch, and how to maintain a feature store for online and offline consistency.

3. Propose Model Architecture

Describe a two-stage system: candidate generation (e.g., embedding-based retrieval) and ranking (e.g., learning-to-rank with contextual features). Include online learning or periodic retraining to adapt to evolving interests.

4. Implement Feedback Loops and Evaluation

Explain how to use implicit feedback to update models, and how to evaluate with online A/B tests and offline metrics. Discuss counterfactual evaluation to avoid bias.

5. Address Trade-offs and Scalability

Discuss trade-offs: model complexity vs. latency, personalization vs. exploration, and how to scale with distributed training and serving. Mention monitoring and drift detection.

Key Points to Mention

Real-time vs. batch processing for user signals
Embedding-based retrieval and approximate nearest neighbor search
Online learning and incremental model updates
Exploration-exploitation balance to avoid filter bubbles
Evaluation metrics: CTR, dwell time, retention, and diversity
Infrastructure: feature store, model serving, and A/B testing

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.