Ashok Kumar

Title of the Talk:
Self-Optimizing AI Infrastructure: From Prediction to Autonomous Control

Abstract :
“Modern infrastructure is exceptionally good at telling us what just went wrong. The next frontier is building systems capable of understanding what is likely to happen next and safely acting before failure occurs. In this keynote, Ashok Kumar explores the evolution from reactive monitoring to predictive, adaptive, and self-optimizing AI infrastructure. Drawing from his 2026 IEEE Access research, Ashok demonstrates how database execution plans can be transformed into learnable representations that predict memory, CPU, and runtime requirements before execution, enabling smarter infrastructure-routing decisions. The evaluated approach reported an 82% reduction in out-of-memory failures and approximately 33% lower infrastructure cost under the study conditions. The discussion then expands beyond resource prediction into two emerging challenges: trace-aware evaluation of long-running AI agents and closed-loop control of distributed cloud infrastructure. These areas are reflected in Ashok’s 2026 U.S. patent application and German utility-model publication. The keynote brings these ideas together through a practical engineering model: Observe → Represent → Predict → Decide → Enforce → Learn. The objective is not uncontrolled AI autonomy. It is infrastructure that is measurable, confidence-aware, policy-governed, reversible, and capable of progressively optimizing its own behavior.”

Bio:
“Ashok Kumar is a Staff Software Engineer at Walmart Global Tech, IEEE Senior Member, researcher, and inventor whose work focuses on building intelligent, reliable, and scalable computing infrastructure.

His engineering experience spans large-scale distributed systems, e-commerce platforms, observability, web performance, cloud infrastructure, cost optimization, and AI-enabled engineering systems. His work has addressed challenges across last-mile delivery infrastructure, marketplace and seller platforms, international e-commerce, cloud-cost governance, and enterprise performance systems.

Ashok is a co-author of the 2026 IEEE Access research “Learning Execution Plan Embeddings for Multi-Dimensional Query Resource Prediction,” which investigates how execution-plan structures, machine learning, similarity search, and graph-based representations can predict computing-resource requirements before workload execution.

His intellectual-property work extends this research direction into long-running AI-agent reliability and autonomous infrastructure control. His U.S. patent application addresses trace-aware evaluation of AI agents across long execution trajectories, while his German utility-model publication describes closed-loop technical control of web-performance parameters across distributed cloud systems.