How do you approach the trade-offs between model latency and predictive accuracy in a production system?
