Whenever you're ready, walk me through how you'd approach this.
How do you balance the trade-offs between latency and accuracy when serving AI models to end users?
Explain how you would design the serving path, choose models or model variants, allocate a latency budget, and decide when to use caching, batching, quantization, or fallbacks. Discuss how you would evaluate the trade-off offline and online, including monitoring for feature drift, training-serving skew, and quality regressions.