A customer wanted an llm system for complex contract question answering tasks. We helped them build it—beating the baseline by 64 points.