YouTube Video
Project Description
This project is an MLOps Incident Commander built on the hackathon starter stack, designed to turn agent reasoning into an operational control surface instead of a plain chat transcript. It monitors scenarios like data drift, latency spikes, and accuracy drops, then presents live incident cards, timelines, and metric visualizations that update as the agent works.
What makes it strong for judging is the combination of originality and execution: the agent does not just explain incidents, it drives a human-in-the-loop remediation flow with actionable approvals, state updates, and runbook progression. The interface reacts to agent output in real time through structured events, so users can see charts, tool activity, and status changes immediately.
The generative UI stack uses CopilotKit with AG-UI event streaming, LangGraph for stateful orchestration, and MCP integration for tool and widget interoperability. Recharts and selective D3 overlays power dynamic dashboards, while webhook-driven incident ingestion enables realistic, interactive, end-to-end MLOps operations beyond text chat.