Design an Image Moderation Service (NSFW Detection)

To design a scalable Image Moderation Service, I would start by defining the core requirement: processing millions of uploads asynchronously without blocking the user experience. The system begins when a user uploads an image to our Object Storage layer. This triggers an event sent to a high-throughput message broker like Kafka, ensuring we can absorb traffic spikes during viral events. A fleet of stateless worker services consumes these messages. We implement a tiered detection strategy: first, a fast heuristic filter removes obviously safe or malicious files; then, a deep learning model analyzes the remaining images for NSFW content. If the model confidence is below a threshold, say 80%, the image is routed to a human review queue rather than being auto-approved or rejected. This ensures safety without false positives ruining user trust. We need a dedicated database schema to track image status, model version used, and reviewer ID. To handle scale, workers auto-scale based on queue depth, and we use Redis to cache recent image hashes to prevent redundant processing. Finally, we must ensure the human review interface prioritizes urgent content, allowing moderators to act quickly on flagged items while the system continues processing the backlog efficiently.