Your LLM feature works in demos and disappoints in production. I review the whole path - prompts, grounding, model choice, cost per call, failure handling - and give you a written findings document with a prioritised fix list and what each fix is worth.