Hemanta Ghosh

Silent Anomaly Detection in Agentic AI Networks Using MSAS

Abstract:

As organizations adopt Agentic AI, AI agents increasingly rely on external tools such as APIs, databases, and microservices to perform real business tasks. These tools often change over time, but the AI agents are not always aware of those changes. As a result, they may continue producing incorrect outputs without any errors or alerts. This talk introduces Multi-Dimensional Silent Anomaly Scoring (MSAS), a lightweight runtime monitoring framework that detects these hidden failures by monitoring schema, semantics, latency, and behavior. The session will explain the silent anomaly problem, demonstrate how MSAS improves the reliability of production AI systems, and discuss a future direction called Behavioral Fingerprinting for continuous AI tool identity verification.

Profile:

I am an Engineering Manager at Best Buy Inc., USA, with over 22 years of experience in software engineering, enterprise architecture, cloud computing, artificial intelligence, and large-scale distributed systems. My work focuses on designing AI-driven enterprise platforms and leading engineering teams to deliver scalable, production-grade solutions. In addition to my industry experience, I actively contribute to the research community through publications, peer review, and invited speaking engagements at international IEEE and Springer conferences. To complement my technical expertise with business leadership, I earned an MBA from the Carlson School of Management, University of Minnesota, strengthening my strategic, financial, and organizational management skills.