How do you balance latency, throughput, and cost when designing systems in a cloud environment for Character.AI?
