Section 9: The Scammer Reads the Same Docs You Do - Deep Dive: Trust & Safety
Executive Summary: A marketplace dies of distrust long before it dies of latency. We sort the threats into four families, then build a layered defense: account friction, synchronous rules inside the 200 ms posting budget, asynchronous classifiers with a p95 of 20 seconds, a publish/hold/reject decision, a human queue sized at ~83 shifts a day instead of ~4,167, and post-publish takedowns that clear search in seconds. Plus stolen-photo hashing, shadow restriction's ethics, chat scanning and a classifier outage.
Section 8 got Daniel's query to the Trek in under 300 ms. The interviewer's next question is the one that matters to Daniel: "How do you know that bike is real?"
You: "I can't know it, but I can make faking it expensive. This adversary adapts, so instead of one smart filter I'll build several cheap layers, each catching what the earlier ones miss, with humans for the uncertain cases."
Know the Enemy: Four Threat Families
| Family | What it looks like | Why it's hard |
|---|---|---|
| Spam and reposts | The same item posted in 30 cities; one listing reposted daily to stay on top | Each copy looks innocent on its own |
| Prohibited items | Weapons, drugs, animals where restricted, recalled goods | Sellers use euphemisms and misspellings |
| Scams | Rental listings with stolen photos; overpayment with a fake check; off-platform payment links; the verification-code scam aimed at sellers like Maya | The listing is often fine; the fraud lives in the chat |
| Harassment | Abusive messages, a buyer who won't take no | It happens after the listing is approved |
A listing-only filter misses most scams: trust and safety has to watch accounts, listings and conversations.
Sign in to continue reading
The rest of this lesson is available with a free account. Signing in with Google or Microsoft is free.
Sign in to read the full lesson