Jyothish Sreedharan
Modern companies struggle to maintain data pipelines that process millions of events per second while meeting strict performance requirements. Traditional methods use fixed configurations and manual fixes, which fail when workloads change unexpectedly. This paper presents an AI-powered system that automatically manages Apache Flink pipelines on Kubernetes. The system uses machine learning to predict problems before they occur, recover from failures automatically, and optimize resource usage continuously. The system was evaluated using two publicly available benchmark datasets: the NYC Taxi Trip Record dataset adapted for streaming scenarios and the Yahoo Cloud Serving Benchmark dataset. Tests show the AI-driven approach significantly reduces service violations, substantially cuts recovery time, and lowers infrastructure costs compared to manual management while maintaining better performance. The system uses three AI agents working together where the prediction agent forecasts problems ahead of time with high accuracy using a neural network that processes multiple metrics continuously, the recovery agent detects failures rapidly using isolation forests, autoencoders, and long short-term memory networks, and the optimization agent adjusts resources dynamically based on workload patterns using reinforcement learning. Together, these agents enable the system to operate autonomously, dramatically reducing manual interventions and operational overhead.