A fraud startup scores payments in under 100 ms with spiky daytime traffic and idle nights, using a custom XGBoost container. It needs automatic scaling, A/B testing of model variants, and minimal ops overhead. Which ML inference architecture using SageMaker AI endpoints meets these needs?

Real-time endpoints with autoscaling and endpoint variants fit because they keep compute warm for synchronous sub-100 ms inference, support custom XGBoost containers, and can scale with invocation or latency metrics during daytime spikes and down at night. Production variants let you shift traffic by weight for A/B testing, while AWS manages endpoint patching and health checks to reduce operational work. Asynchronous endpoints with autoscaling and custom containers suit queued scoring and large payloads, but requests wait for background processing, so queueing can violate the latency target. Serverless inference handles intermittent traffic and idle periods with pay-per-use billing, but cold starts after inactivity threaten sub-100 ms scoring; its concurrency-based scaling also does not provide the same metric-driven autoscaling and variant-weight traffic shifting for A/B tests. Real-time endpoints with variants and scheduled daytime capacity preserve low latency and A/B testing, but a fixed schedule cannot react to metric-driven spikes, risking throttling or wasted capacity compared with target-tracking autoscaling. Exam caveat: first match the latency and interaction pattern—synchronous, queued, or serverless—before comparing scaling and cost. Practical check: load-test p99 latency at target throughput, verify scaling policies and cooldowns, and confirm variant weights, CloudWatch alarms, and rollback behavior in staging.