uniopen fine-tunes Amazon Nova for retail moderation

uniopen, Taiwan's Uni-President Enterprises Group digital platform, adapted Amazon Nova 2 Lite to its custom content moderation policy using supervised fine-tuning in Amazon SageMaker AI. The policy classifies interactions across nine behavior categories and three subject types. A human-verified feedback loop converts reported errors into training data stored in Amazon S3, while Argo Workflows on Amazon EKS orchestrates evaluation and deployment. Hard gates block automatic promotion on regression failures; soft gates flag low-confidence results for administrator review. The fine-tuning dataset contained 3,391 training windows, evaluated against a held-out set of 737 conversation windows.