Multi-Agent AI Framework for Outlier Detection

Objective:

A next-generation anomaly detection framework powered by a Multi-Agent AI architecture. This system uses intelligent agents that work in parallel to monitor, classify, and escalate anomalies across business metrics enabling near real-time observability and intelligent incident triage.

This project introduces a multi-agent orchestration model to::

• Reduce manual monitoring and investigation
• Detect issues early with high accuracy
• Proactively notify stakeholders to prevent revenue and productivity loss

Outlier Detection Logic :

• Threshold Calculation:

  • For each segment and metric, compute the Monthly Daily Average over the last 6 completed months

  • Define thresholds as:

    -->Upper Threshold = 200% of highest historical daily average maximum

    -->Lower Threshold = 50% of lowest historical daily average minimum

• Detection Window:

  • Apply the thresholds to the most recent 15-day window
  • Detect anomalies across all segments and metrics

Architecture

Agent Roles & Responsibilities

• Supervisor Agent :
  • Co-ordinates and monitors the execution of all agents.
• Stats Agent :
  • Cleans data, calculates monthly daily averages, and defines thresholds per segment.
• Analyst Agent :
  • Processes recent data to compute daily averages and compares them against thresholds.
• File Checker Agent:
  • Verifies the completion of Stats and Analyst tasks and reports back to the Supervisor Agent.

Tech Stack:

• Languages: Python
• Frameworks: LangGraph, LangChain
• Infrastructure: Airflow

Outcomes

• Hands-on experience with LangGraph and LangChain and its architecture, including agent state management.
• Effective for non-critical production applications.
• Inspired the creation of an AI-powered incident management system for Data Engineers:
  • Achieved 30% improvement in engineering productivity.
  • Reduced manual debugging and resolution time for production failures

Next Steps

• Integrate Agentic RAG (Retrieval-Augmented Generation) for contextual decisioning.
• Explore managed service deployment with observability and guardrails.

How to Run (Local Simulation)

git clone https://github.com/krishnamami/Multi_Agent_Anamoly_Detection.git cd Multi_Agent_Anamoly_Detection python agents.py

Related Projects

Fine_Tuning_LLM

Markov_Chain_Attribution

Distributed ML SageMaker Pipeline

Author

Krishna Goud

Head of Data Engineering & MLOps | Rocket LA LinkedIn

Delivering $4B+ business impact via AI-first, scalable, real-time data systems