How would you optimize an LLM application at CORAS that is currently experiencing high latency and high API costs?