Skip to content
Kml
All projects
Personal Project

Incident Root-Cause Copilot

Multi-agent AIOps system designed to correlate anomalies, recent changes, and historical incidents to support root-cause analysis.

  • FastAPI
  • React
  • Docker
  • Observability
Incident Root-Cause Copilot landing page with an example incident: a P95 latency spike chart for a checkout service, an AI confidence score, and a link to the recommended remediation.

Overview

Root Cause is a personal project: an AI incident copilot for SRE teams. It detects anomalies, correlates them with recent deployments, searches past incidents, and recommends a fix. Specialised agents each investigate one source of evidence, and their findings are combined into a root-cause hypothesis with a confidence score.

The Problem

During an incident, engineers must manually connect metric anomalies, recent deployments or configuration changes, and knowledge from past incidents. That context switching takes time when it matters most.

Architecture

  1. Signals

    • Service metrics
    • Deployments & changes
    • Incident history

  2. Agents

    • Anomaly detection
    • Change correlation
    • Incident history search

  3. Analysis

    • Root-cause hypothesis
    • Confidence score
    • Recommended remediation

  4. Interface

    • React incident view
FastAPI backend; all services run in Docker containers.

How It Works

  1. 1

    Detect

    Anomalies, such as a latency spike in a service, are detected from metrics.

  2. 2

    Correlate

    Agents check deployments and changes made around the time of the anomaly.

  3. 3

    Search history

    Similar past incidents are retrieved to add context.

  4. 4

    Recommend

    A root-cause hypothesis with a confidence score and suggested remediation is shown.

My Contribution

  • Designed and built the project end to end.
  • Developed the multi-agent backend with FastAPI.
  • Built the React interface for exploring incidents and correlated evidence.
  • Containerised the services with Docker.

Engineering Challenges

  • Splitting the investigation into agents with clear, non-overlapping responsibilities.
  • Presenting each hypothesis with its evidence and a confidence score, so the output stays verifiable.

Lessons Learned

  • Agent systems are easier to trust when every conclusion links back to its source signals.