How do you balance model complexity with latency requirements when designing real-time inference pipelines?
