How uniopen customized Amazon Nova to their retail moderation policies for production deployment

uniopen, a digital communication and membership platform launched by Taiwan's Uni-President Enterprises Group, adapted Amazon Nova 2 Lite to its business-specific retail moderation policies, according to an AWS Machine Learning Blog post. The moderation policy classifies each interaction along two axes: what behavior occurred, across nine categories, and what subject the behavior refers to, whether brand, other, or forbidden. Both must be correct for a moderation decision to be useful.

The customization used supervised fine-tuning with Low-Rank Adaptation (LoRA) in Amazon SageMaker AI, followed by a prompt-level output optimization. The prompt change simplified model output from JSON to a line-based format and clarified how multiple behaviors should be returned; it required no additional model training.

All three configurations were evaluated on the same held-out test set of 737 conversation windows. The fine-tuning dataset contained 3,391 training windows. The baseline Amazon Nova 2 Lite scored a Per Behavior Macro F1 of 0.5852 and a Subject Type Macro F1 of 0.4162. After fine-tuning, the scores rose to 0.8364 and 0.8302. After prompt optimization, they reached 0.8550 and 0.8491, exceeding production targets of 0.8500 and 0.8200 respectively.

The architecture separates the production moderation path from correction, training, evaluation, and deployment. Amazon Nova 2 Lite handles primary moderation requests, while Amazon Nova 2 Pro supports candidate correction generation for reported errors, with a human reviewer required to verify each correction before it enters the training set. Amazon S3 stores the verified correction set and training data, DynamoDB tracks active and candidate model configurations, Argo Workflows on Amazon EKS orchestrates prompt optimization, evaluation, and deployment, and Argo CD applies approved configurations to production. Amazon SNS and Amazon CloudWatch notify operators when a hard gate fails or a candidate needs attention.

Two kinds of checks control promotion. Hard gates are must-pass regression tests; a failure stops the workflow and sends an alert. Soft gates are warning signals such as low confidence or a drop in performance for a specific class; a soft-gate warning keeps a candidate pending administrator review and approval before deployment. Human review remains mandatory for ambiguous cases and for corrections reused as training data. The team will continue using both metrics as production gates and collect new boundary cases from real traffic. Model availability varies by AWS Region.