Download to read offline
![LLM PROMPT FLOW
System Message
“You are an AI-driven RCA engine. Analyze logs,
metrics, traces, and alerts to identify the most
probable root cause, summarize impact, and
recommend fixes. Keep output concise and evidence-
based.”
User Message (Telemetry Input)
{
"incident_id": "INC-001",
"logs": [
{"svc": "checkout", "msg": "inventory lookup failed"},
{"svc": "inventory", "msg": "db timeout"}
]
}
LLM
Generate
RCA
Prompting
• Zero-shot
• Few-Shot
• Chain-Of-Thought(COT)](https://image.slidesharecdn.com/kamalbisht-251127145431-bde77bf7/75/OSMC-2025-Generative-AI-for-Observability-Automating-Root-Cause-Analysis-in-Modern-IT-Systems-by-Kamal-Bisht-pdf-16-2048.jpg)
Uploaded byNETWAYS
OSMC 2025: Generative AI for Observability: Automating Root Cause Analysis in Modern IT Systems by Kamal Bisht.pdf
As modern IT systems grow in complexity, the traditional approaches to root cause analysis (RCA) are proving too slow, too manual, and too reactive. Observability platforms produce a flood of logs, metrics, and traces—but transforming that raw telemetry into actionable insights remains a major challenge. This session explores how Generative AI can be used to automate RCA by understanding system behavior, correlating disparate signals, and generating human-readable incident reports. By leveraging pre-trained language models and retrieval-augmented architectures, AI agents can now analyze incidents contextually and assist engineers in identifying probable root causes—dramatically reducing mean time to resolution (MTTR). Key takeaways: How to design observability workflows enhanced with generative reasoning. Techniques for summarizing, correlating, and interpreting log and metric data using AI. Real-world architecture for integrating Generative AI into operational pipelines (e.g., prompt templates, embeddings, RAG).
More Related Content
Similar to OSMC 2025: Generative AI for Observability: Automating Root Cause Analysis in Modern IT Systems by Kamal Bisht.pdf
OSMC 2025: Generative AI for Observability: Automating Root Cause Analysis in Modern IT Systems by Kamal Bisht.pdf
- 1. ©2025 Discover, adivision of Capital One, N.A. Opinions are those of the individual author. Unless noted otherwise in this post, Discover is not affiliated with, nor endorsed by, any of the companies mentioned. All trademarks and other intellectual property used or displayed are property of their respective owners
- 2. HI I’M KAMALSINGH BISHT Observability & AI/ML Technologist | IEEE Senior Member | Author & Speaker ▪ A seasoned technologist with extensive experience spanning observability, AI/ML, and cybersecurity. ▪ Currently at Discover, with prior senior engineering roles at JPMorgan Chase and Zillow. ▪ IEEE Senior Member, conference speaker, and author of multiple AI-driven observability publications. https://www.linkedin.com/in/kmluvce/
- 3. AGENDA • Overview • Challenges •Architecture • RCA Workflow • Improvements • Takeaways
- 4. THE STATE OFMODERN IT SYSTEMS • Explosion in telemetry data (logs, metrics, traces) • Complex microservices with distributed failure domains • Manual Root Cause Analysis (RCA) still dominates incident workflows • High Mean Time to Detect and Mean Time to Resolve
- 5. WHY TRADITIONAL AIOPSFALLS SHORT
- 6. ARCHITECTURE OVERVIEW Dynamic &Adaptive Behavior – 1 2 3 4
- 7. DATA SOURCE Data Source Monitoring Tools GenAI-RCA GenAI- RCA Type 1 Type 2 Alerts Logs/ Metrics/ Traces
- 8. DATA SOURCE Data Source Monitoring Tools GenAI-RCA GenAI- RCA Type 3: Hybrid Alerts Logs/ Metrics/ Traces
- 9.
- 10.
- 11.
- 12.
- 13. RCA ENGINE(RAG –INGESTION)
- 14.
- 15.
- 16. LLM PROMPT FLOW SystemMessage “You are an AI-driven RCA engine. Analyze logs, metrics, traces, and alerts to identify the most probable root cause, summarize impact, and recommend fixes. Keep output concise and evidence- based.” User Message (Telemetry Input) { "incident_id": "INC-001", "logs": [ {"svc": "checkout", "msg": "inventory lookup failed"}, {"svc": "inventory", "msg": "db timeout"} ] } LLM Generate RCA Prompting • Zero-shot • Few-Shot • Chain-Of-Thought(COT)
- 17.
- 18. LLM COMPARISON Model TypeKey Strengths Limitations / Considerations GPT-4 Turbo (OpenAI) Proprietary (Closed- Source, API- based) Proven reasoning accuracy, reliable RCA summaries Cost, black-box, Data privacy & compliance Llama 3 (Meta) Open-source Strong generalization, fine-tunable GPU-intensive Mistral (mistral AI) Open-source lightweight, fast, cost- efficient, on-prem ready Limited reasoning depth Sources: OpenAI (GPT-4 Turbo Docs, 2025); Meta AI (Llama 3 Model Card, 2025); Mistral AI (Official Model Documentation, 2025); Author’s empirical evaluation, 2025.
- 19. TECHNICAL & ARCHITECTURALLIMITATIONS Limitation Mitigation Model Hallucination RAG integration Context Window Sliding memory buffers Compute Demand Vector caching Integration Modular API design
- 20. RESOURCES GitHub: https://github.com/rootiq-ai Published Papers: • GenerativeAI–Driven Observability for Automated Root Cause Analysis in Modern IT Systems: Architecture and Vision • Convergence of AI and Observability: Predictive Insights Automation in Modern IT Operations Hugging Face • https://huggingface.co/ LangChain • https://www.langchain.com/ LangGraph • https://www.langchain.com/langgraph
- 21. Dynamic & AdaptiveBehavior –