E-commerce platforms know what their customers bought, but raw purchase logs are hard to reason about — for humans and for LLMs alike. This work turns transaction histories into natural-language customer personas: compact, interpretable profiles that describe what kind of shopper a customer is.

The approach

Rather than prompting an LLM with a raw item list, the framework distills purchase records into structured persona descriptions that capture preference dimensions such as price sensitivity, category affinity, and brand loyalty. The personas are generated at scale with cost-efficient LLM pipelines and validated for faithfulness against the underlying behavior.

What personas buy you

Personas serve as a human-readable interface between behavioral data and applications: customer segmentation, cold-start recommendation, marketing copy targeting, and any downstream task where an LLM benefits from knowing who the user is. Experiments on real e-commerce data show persona-augmented pipelines outperform purchase-history prompting on both quality and token cost.