The real cost driver was reprocessing, not inference. Two cents after running this through prod traffic. Design the rollout so rollback takes minutes, not hours.

Keep the system operable by the smallest competent team. If volume is spiky, budget for burst capacity explicitly. Biggest risk is invisible complexity creeping in.

A thin rules layer filtered 80% of noisy alerts. Small anecdote that might help. Ship small and often to limit blast radius.

We phased traffic: 5% → 25% → 60% → 100% to de-risk. Design the rollout so rollback takes minutes, not hours. We learned this the hard way during a vendor swap.

Latency budgets forced us to drop nice-to-haves. Two cents after running this through prod traffic. Optimize observability before model tuning.