Cut generative inference spend 58% ($2.1M annualized) by introducing Model-Tiering (routing 78% of traffic to smaller models via a learned classifier), prompt caching, and reserved provisioned throughput for peak windows — without measurable regression on the eval suite. Designed and led delivery of a Bedrock-based agentic support platform — tool-calling agents over a curated skill library with SSE streaming — that now handles 40M+ requests/month and deflects 61% of tier-1 support tickets, reducing median time-to-resolution from 34 minutes to 7.