From "Who Is This User?" to "What Does This Purchase Mean?": A Deployed Pipeline for Semantic User Profiling at Bank Scale

A deployed pipeline reduces semantic profiling costs by performing LLM inference on transaction patterns rather than individual users. This three-phase approach—resolving item names, profiling patterns, and tagging clusters—maintains AUC parity with per-user inference while decoupling the compute budget from total user count.