Stream: Virtual Room 4
Time: 11:15 - 12:00
Mainframe abends are still one of the toughest operational issues in enterprise computing. Even after many years of investing in tools, the analysis process is mostly manual. When an abend occurs, an operator reviews the dump, checks documentation, looks at recent changes, and tries to fix the problem. What if an AI assistant could handle all of that on it’s own, in a conversation, and within seconds? This presentation introduces a new setup that connects Abends telemetry data with Large Language Models (LLMs) using the Model Context Protocol (MCP). This is a new open standard for integrating AI tools. We built a special MCP server that offers 14 dedicated analytical tools that an LLM can use to examine mainframe program failures, evaluate stability, spot anomalies, and link code changes to production issues. Using actual production data from 12 mainframe LPARs, we show how this system: - Automatically classifies and sorts over 1,900 abends across 156 programs with a composite risk score - Detects spikes in anomalies, like 5.6 times increase from 79 to 441 daily abends, using statistical baselines - Reveals that 97% of failing programs had no recent code changes, challenging the common belief that "recent changes cause failures" - Classifies programs into health categories (Healthy Evolution, Legacy Burden, Stable Asset) based on abend frequency, code changes, and debugging efforts - Calculates Mean Time Between Failures (MTBF), ranging from 2 hours (critical) to over 270 hours (stable) across the program set We will show a live demonstration of an LLM performing a multi-step investigation. This will include summarizing the initial abend, analyzing abend code, assessing program stability, and providing actionable recommendations — all through natural language conversation. This work builds on IBM's Orthogonal Defect Classification (ODC) and applies defect tracking ideas from Fredericks & Basili (1998) to the AI age. In this era, classification and analysis are done by an AI assistant using structured tools instead of manual form-filling.
There is currently no attachment for From Abend to Insight: How LLMs and MCP Server Tools Are Transforming Mainframe Defect Detection
Krutika Sapkal is a Data Scientist at BMC Software with over 8 years of experience in Data Science, Data Engineering, and cloud-based analytics solutions. She specializes in building scalable data platforms and machine learning systems using technologies such as Python, Azure, AWS, Databricks, and Elasticsearch. Throughout her career, she has led large-scale data migrations, designed secure and cost-efficient data architectures, and deployed machine learning models using MLOps and Kubeflow. Krutika also holds one published patent and six international research publications in areas including NLP, parallel computing, and data analytics. She is passionate about leveraging AI and data-driven insights to improve system reliability, development efficiency, and business decision-making.
Click here to give some Feedback so we can make it even better next year!