Step 1: use a different LLM that isn't 10x slower and 2-5x more expensive for the same level of quality.