Ashok Kumar

Beyond the Final Answer: Engineering Reliable Agentic AI for Real-World Systems

Abstract:

Agentic AI is rapidly evolving from conversational systems into autonomous software capable of reasoning, using tools, maintaining context, and executing long-running workflows. As these systems become more capable, a critical engineering challenge emerges: how do we determine whether an AI agent behaved reliably across an entire execution journey—not simply whether its final answer appears correct?

This keynote explores the engineering foundations required to move agentic AI from impressive demonstrations to reliable real-world systems.

The talk will examine challenges including long-running execution, context preservation, reasoning and action traces, tool interactions, decision dependencies, failure localization, confidence-aware automation, and responsible autonomy.

Drawing from my research and intellectual-property work, I will discuss the concept of trace-aware AI evaluation, where an agent is treated as a temporal execution trajectory rather than a single input-output interaction. My 2026 U.S. patent application, US-20260186944-A1, addresses this problem through an architecture that preserves contextual state, reasoning/action sequences, temporal execution records, and decision dependencies for evaluating long-running AI agents.

The keynote will organize these ideas around a practical lifecycle:

Observe → Trace → Evaluate → Detect → Govern → Act → Learn

The central message is that the future of applied AI will depend not only on building more capable models, but on engineering systems in which autonomy is observable, measurable, explainable, policy-governed, and reliable enough for real-world deployment.

Profile:

Ashok Kumar is a Staff Software Engineer at Walmart Global Tech, IEEE Senior Member, researcher, and inventor whose work spans distributed systems, large-scale e-commerce platforms, web performance, observability, cloud optimization, and intelligent infrastructure. His engineering experience includes marketplace and seller systems, international e-commerce, performance governance, observability modernization, and cloud-cost optimization.

His research and inventions focus on advancing computing systems from reactive monitoring toward predictive, adaptive, and autonomously controlled infrastructure.

He is a co-author of the 2026 IEEE Access paper “Learning Execution Plan Embeddings for Multi-Dimensional Query Resource Prediction,” which explores execution-plan representations, machine learning, similarity search, and graph-based models for predicting memory, CPU, and runtime requirements before execution.

Ashok is also a named inventor on the U.S. patent application US-20260186944-A1, addressing trace-aware evaluation of long-running AI agents by preserving temporal execution context, reasoning/action sequences, and decision dependencies.

His German utility model DE202026100134U1 extends performance engineering beyond passive monitoring by describing a closed-loop architecture that collects distributed web-performance state, generates control signals, and adjusts system resources or execution parameters.

Together, his papers, inventions, and production engineering work represent a consistent technical contribution: building systems that can observe complex behavior, predict resource and reliability risks, evaluate outcomes, and take bounded actions to improve performance and efficiency.