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

SeniorPrefer not to say

Jun 2026

Summary

System design round at Bytedance for a software engineer role, focused entirely on building a content moderation platform for a short-video product. The scope was massive and I don't think I fully covered everything they were looking for, but it was a genuinely interesting problem.

Questions Asked (1)

Design a large-scale content moderation system for a short-video platform that handles video, captions, comments, audio, and metadata, and can detect violations like spam, nudity, violence, hate speech, self-harm, and copyright issues, with both pre-publish and post-publish checks.

System DesignTechnical Trade-offsProduct Analytics & Metrics

Author's notes

This is a beast of a question.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying requirements and scale, then propose a multi-stage pipeline that combines automated ML classifiers with human review, covering both pre-publish and post-publish checks. Emphasize trade-offs between latency, accuracy, and cost, and describe how to handle different content types and violation categories.

Pro tip: Show awareness of adversarial evasion and the need for continuous model retraining; mention feedback loops from human reviewers and user appeals to improve detection over time.

1. Clarify Requirements and Scale

Ask about daily upload volume, peak QPS, latency SLAs, types of violations, and regional policies. Establish assumptions for scale (e.g., millions of videos/day) and prioritize pre-publish vs post-publish checks.

2. Design Multi-Stage Moderation Pipeline

Propose a pipeline: fast automated checks (hash matching, lightweight classifiers) for immediate decisions, followed by deeper ML models for nuanced violations, and human review for edge cases. Include both pre-publish (blocking) and post-publish (async) flows.

3. Detail Component Architecture

Describe services for ingesting content, extracting features (video frames, audio transcripts, text), running classifiers (e.g., CNN for nudity, NLP for hate speech), and aggregating scores. Include storage for moderation results and a queue for human review.

4. Address Scalability and Reliability

Discuss horizontal scaling of ML inference (GPU clusters), caching, sharding by content ID, and handling failures with retries and dead-letter queues. Ensure low latency for pre-publish and high throughput for post-publish.

5. Cover Trade-offs and Metrics

Explain trade-offs: precision vs recall, latency vs accuracy, cost of human review. Define metrics: violation detection rate, false positive rate, review latency, and user appeal rate. Mention A/B testing for model improvements.

Key Points to Mention

Multi-modal detection: separate models for video, audio, text, and metadata, with fusion of signals.
Pre-publish vs post-publish: pre-publish for high-risk violations (e.g., nudity) with low latency, post-publish for nuanced cases (e.g., hate speech) with higher accuracy.
Human-in-the-loop: escalation to human reviewers for ambiguous cases, with feedback to retrain models.
Scalability: use of distributed systems, GPU inference, and message queues to handle high volume.
Adversarial evasion: techniques like perceptual hashing, ensemble models, and continuous monitoring.
Metrics and monitoring: track false positives/negatives, review time, and appeal rates to optimize system.

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