Your question is Low-Latency Inference Patterns. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
What architectural patterns do you prefer for low-latency inference in trading systems?
Discuss how you would separate model training from inference, choose between in-process, service-based, and hardware-accelerated serving, and preserve correctness under strict latency and availability requirements. Address deployment, data movement, model updates, observability, and degraded-mode behavior.