Microsoft·Software Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say

Jun 2026

Summary

Microsoft system design round, got asked to build a home recommendation feature for a real estate platform. Pretty open-ended and I wasn't totally sure how deep to go on the ML side versus the infrastructure side.

Questions Asked (1)

Design a system that recommends similar homes for sale on a real estate platform like Zillow.

System DesignProduct Sense & IdeationTechnical Trade-offs

Author's notes

I started with clarifying questions about what 'similar' means, which felt smart at the time, but I think I spent too long on that part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying requirements and scale, then outline a high-level architecture that separates offline model training from online serving. Focus on feature engineering for home similarity, candidate generation, and ranking, and discuss trade-offs between different similarity metrics and scalability approaches.

Pro tip: Emphasize how you would handle cold-start for new listings and incorporate user feedback loops to continuously improve recommendations, showing product sense beyond pure technical design.

1. Clarify Requirements and Scope

Ask questions to understand scale (number of homes, users, QPS), latency requirements, and what 'similar' means (location, price, features, photos). Define success metrics like click-through rate or user engagement.

2. High-Level Architecture

Propose a two-phase system: offline pipeline for feature extraction and model training, and online serving for real-time recommendations. Include components like data ingestion, feature store, candidate generation, ranking, and caching.

3. Feature Engineering and Similarity

Discuss features: structured (price, beds, baths, sqft, location), unstructured (text descriptions, images). Explain similarity metrics (cosine, Euclidean) and how to combine them, possibly using embeddings from neural networks.

4. Candidate Generation and Ranking

Describe how to efficiently retrieve top-N similar homes (e.g., using approximate nearest neighbor search like FAISS, or geo-based indexing). Then rank candidates using a machine learning model that incorporates user behavior and context.

5. Scalability, Trade-offs, and Evaluation

Address scaling to millions of listings (sharding, distributed processing), trade-offs between accuracy and latency, and offline/online evaluation (A/B testing, metrics). Mention handling cold-start and feedback loops.

Key Points to Mention

Use of embeddings (e.g., from home images and text) to capture semantic similarity beyond structured features.
Approximate nearest neighbor (ANN) search for efficient candidate retrieval at scale.
Two-tower neural network architecture for candidate generation and ranking.
Feature store for consistent online/offline features and low-latency serving.
Handling cold-start for new listings by using content-based features and fallback to popularity or location-based recommendations.
A/B testing framework and online metrics (CTR, conversion) to evaluate and iterate on the recommender.

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