Velocity rules at authorization

Design the velocity-rules engine that runs inline in card authorization and blocks patterns like 'this card was tried >5 times in 60s', 'this device touched >10 cards in 1h', 'this BIN spiked 50x above baseline'. ~25,000 auths/sec, decision budget <15ms, and the counts must be accurate enough that a real fraud burst is caught within seconds, while a legitimate retail spike doesn't false-positive everyone. How do you maintain millions of sliding-window counters at this rate and latency?