Sylendran Arunagiri

This talk argues that building effective AI agents relies on data flywheels rather than simply using the largest available language models. A data flywheel is a continuous cycle of data processing, model customization, evaluation, and safety guardrailing. This process allows agents to refine their performance over time by learning from user feedback and production data, ultimately enabling the use of smaller, more cost-effective models without sacrificing accuracy.