Bot or human? An ad platform gets 5 milliseconds to decide.

transcript

An ad request arrives. Is it a bot? You get 5 milliseconds to decide. 500,000 times a second. And every wrong block is lost revenue.

A real-time bid has 100 milliseconds, end to end. Fraud gets under 5 of them, inside the edge budget. At 500,000 inferences a second. So the heavy model can't sit here.

So, tiers. Tier 1 is rules. Known bad IPs, data centre ranges, impossible geography. Under half a millisecond, it blocks about 5 percent. The obvious bots. Tier 2 is a light gradient boosting model, on features cached in Redis. Under 2 milliseconds. Tier 3 is a deep neural network, on the event stream. Off the request path. It updates the blocklists and features.

Tier 2 scores fraud from 0 to 1. Over 0.95? Block. About 2 percent more. But here's the twist. Between 0.5 and 0.95, it doesn't block. It bids lower. Under 0.5, allow.

Why not just block the middle? Because false positives cost revenue. Bidding lower balances fraud losses against lost revenue.

And it all runs at the edge. Before 5 to 10 bidder calls. Before any budget is reserved. Bad traffic is dropped before it costs anything.

Rules first. A light model next. The heavy model off the path. Cheap checks inline. Deep checks async.

Full breakdown in the lecture.