This article details an automated system for evaluating and improving Amazon product page content using a multi-model fusion approach. By rendering pages as single snapshots, the system employs Vision-Language Models, OCR readability signals, and CatBoost classifiers to provide holistic quality assessments and actionable improvement recommendations via SHAP explanations.
Key points
Holistic page evaluation can be achieved by rendering complete product layouts as snapshots for vision-language models.
Fusing VLM semantic scores with OCR readability signals improves detection of visual and textual quality issues.
SHAP values translate complex model predictions into interpretable, actionable improvement recommendations for users.
Decoupling the frozen VLM backbone from a lightweight CatBoost classifier allows for efficient, domain-specific model calibration.
End-to-end pipelines on SageMaker can effectively scale evaluation across diverse product categories using hybrid modeling.