A research paper presents a deployed pipeline for semantic user profiling using LLM inference, reducing the inference budget from user-count to transaction-pattern count, achieving a three-order-of-magnitude reduction in LLM inference targets in a major Japanese bank's deployment.