A session on building more effective AI agents through model customization, focused on how teams can move from promising prototypes to dependable production systems. The talk explored why general purpose models often fall short on domain specific tasks and showed how customization techniques such as fine tuning, distillation, direct preference optimization, and continued pretraining can improve accuracy, consistency, tool use, latency, and cost. We also covered how to customize Amazon Nova models with Amazon Bedrock and SageMaker AI, evaluate agent quality with both traditional and LLM based methods, and deploy production ready agents using the broader AWS stack.

Slides