Generate, Evaluate, Improve: Agentic Multimodal Content Personalization with Synthetic Evaluation and Closed-Loop Learning for E-Commerce

Abstract

E-commerce personalization has largely optimized which item to recommend while leaving how the item is presented - image, copy, video, and narrative framing - comparatively static. This paper presents an end-to-end agentic system for multimodal advertising creative personalization. The system maps a customer-product pair to a natural-language story, generates story-conditioned image, copy, and short-video variants, evaluates them with synthetic shopper agents before launch, and then improves both the generation policy and the evaluator using live experiment outcomes. The generation layer combines LoRA-adapted diffusion, safe-zone-aware product placement, device-first outpainting, asymmetric shadow synthesis, and a dedicated semantic restoration pass for high-frequency product regions such as text, logos, buttons, and display content. We formulate the serving layer as an advertising mixtureof- experts (Ad-MoE): story experts, modality experts, product-fidelity experts, and evaluator experts are selected by a learned gate and composed through low-rank adapters. The evaluation layer uses a vision-language-model judge as a quality gate and a 100K+ persona synthetic randomized controlled trial as a pre-launch screening layer. The agent evaluator estimates paired purchase-propensity deltas, posterior probability of positive response, stability across repeated simulations, and artifact-induced concordance failures before any live traffic is spent. Finally, the system closes the loop: contextual bandits update story selection, hierarchical Bayesian calibration shrinks synthetic predictions toward realized live effects, and Thompson Sampling rotates creative variants to balance exploitation, exploration, and freshness. In production, the system contributed more than $200M in attributed revenue, improved click-through rate by 8–12%, reduced variant production time from roughly 21 hours to 12 minutes, saved more than 5,400 annual design hours, lowered creative spend by 80%, and expanded testable catalog coverage from 9% to 45%.