You are designing a safety system for a large consumer marketplace that should prevent weapon listings from being suggested to buyers. Listings include title, description, seller-provided category, images, price, and location, and they can appear in search, recommendations, notifications, and related-item modules. Policy requires very high recall on prohibited weapons, but the business also wants to minimize false positives on allowed items such as kitchen tools, sporting goods, and collectibles. The filter must work both when a listing is created and when items are retrieved for downstream ranking surfaces.
| Signal | Value |
|---|---|
| DAU | 85M |
| Active listings | 120M |
| New or edited listings/day | 9M |
| Peak recommendation/search QPS touching listing eligibility | 450K |
| Images per listing | 1.8 avg |
| p99 online eligibility latency budget | 35ms |
How would you design this end-to-end system so prohibited weapon listings are detected and filtered before they can be recommended, while still supporting low-latency retrieval and ranking at marketplace scale?