I Built a Self-Hosted AI Incident Diagnosis Tool That Only Returns a Root Cause When Multiple Diagnoses Agree

k1ngalph0x — Tue, 14 Jul 2026 18:29:28 +0000

Most AI incident diagnosis tools will happily produce a root cause even when the evidence is weak. Argus takes a different approach.

When an anomaly fires, Argus runs five independent diagnoses against the same incident window. If they converge on the same root cause, it returns a confident diagnosis. If they don't, it returns novel instead of pretending it knows the answer.

It's a single Go binary. The first version had Kafka, microservices, and two databases. It looked impressive on paper, but nobody would actually run it. I tore it down into a single process and replaced Kafka with an in-process event bus. Run it with docker run, bring your own Anthropic API key, and your telemetry never leaves the box. It ingests OTLP or Prometheus remote_write; point your telemetry to a single endpoint.

I've validated it on synthetic cases, reconstructed real postmortems (Cloudflare 2019/2022), and my own distributed system. It hasn't yet been tested against messy real-world production telemetry, which is exactly the kind of feedback I'm looking for.

GitHub: https://github.com/k1ngalph0x/argus

I'd genuinely appreciate people trying it out and telling me where the design falls apart, what feels over-engineered, or what you'd change.

DEV Community: k1ngalph0x

I Built a Self-Hosted AI Incident Diagnosis Tool That Only Returns a Root Cause When Multiple Diagnoses Agree