Serving 500 concurrent LLM chats on one 4-core box with tier-aware queueing

A shared LLM backend experienced high latency due to a naive concurrency limit, causing paying users to starve during peak traffic. A Redis-backed tier-aware slot manager was implemented to prioritize paying users, reducing p99 latency from 20+ seconds to under 2 seconds. The solution is ~40 lines of code and prioritizes fairness over FIFO.