Korvus is a proof of concept for AI-driven root cause analysis that you can build and run yourself. A pluggable harness gives an LLM tool-mediated access to your telemetry, with model choice and inference location left entirely up to you. Alertmanager or a human triggers it, producing a structured report for an on-call SRE under time pressure: what happened, blast radius, and ranked root causes with concrete fixes.
Korvus is a proof of concept that an SRE team can design, build, and operate an AI-driven root cause analysis system using free and open source components end to end: Kubernetes for orchestration, an OpenTelemetry-compatible data path as input, and Alertmanager as the trigger. The same approach reaches below the cluster. The demo also traces faults into the underlying infrastructure layer through OpenNebula, showing that root causes in the base layer, not just the application or cluster layer, are within scope. The harness itself is model-agnostic and works equally with open-weight models such as Mistral, Llama, Qwen, or DeepSeek and with proprietary ones, with model choice and inference location left entirely up to the operator.
The one hard requirement across every layer is GitOps over ClickOps. Because infrastructure and application state are versioned in git rather than changed by hand, the agent can navigate the relevant repositories directly and correlate a telemetry anomaly with the specific change that rolled it out, regardless of which layer it occurred at. The talk’s core argument is methodological: with clean telemetry and a clean, git-driven deployment process in place, the entire pipeline, from data access to inference to report generation, is something a team can implement, inspect, and run themselves, on infrastructure they control.
