Our inference bill keeps growing with usage. We've done the obvious (caching, smaller models for easy queries, batching) but I'm looking for the next level: speculative decoding, request routing across model tiers, quantization without quality cliffs. What's in your cost-optimization playbook?