Observe
Collect pods, events, current and previous logs, and the Deployment configuration.
For engineers on the incident clock
From incident
to verified fix.
Investigate Kubernetes failures, reason over the evidence, and review a tested remediation before a human approves a production change.
Start with OOMKilled. Follow the evidence through every step.
The incident investigation tax
The evidence is scattered.
The clock keeps running.
Pods restart. Events scroll past. Previous logs disappear. Even a small configuration failure takes time to reconstruct before a fix can be trusted.
IncidentZero brings the investigation into one traceable workflow: evidence, reasoning, a concrete change, and validation history. You can inspect every decision and take the RCA with you.
One bounded investigation
Evidence in. Reviewable change out.
Reasoning is powered by Nemotron. Research is conditional. Every proposed patch must pass the same deterministic safety policy.
Reason
NVIDIA Nemotron forms evidence-backed hypotheses through Nebius Token Factory.
Research
Consult focused Tavily sources when the reasoning step needs external evidence.
Repair
Generate a concrete YAML patch and enforce a narrow configuration allow-list.
Validate
Evaluate the candidate in the selected environment, with at most two remediation attempts.
Built to make the boundaries visible
A small, inspectable system.
A persisted agent workflow connects the operator experience to model reasoning and controlled validation.
Next.js workspace
Incidents · agent timeline · diff · RCA
FastAPI + LangGraph
Typed boundaries · SQLite · SSE
Nemotron + tools
Nebius Token Factory · Tavily · sandbox
Local Kubernetes mode
Real Kind evidence, policy-approved patches, rollout, owned-pod checks, logs, and HTTP health validation. Production contexts stay read-only.
Hosted simulation mode
Infrastructure evidence and configuration checks are explicitly simulated. Nemotron still generates the diagnosis and patch. Simulated passes are never counted as Kubernetes-verified fixes.
Autonomy with an engineering boundary
Trust comes from
what you can inspect.
Review the collected evidence, model diagnosis, focused sources, exact patch, and named validation checks. Download the report and decide what goes to production.
Open the workspace →- Production remains read-onlyLocal Kubernetes writes target only the fixed demo sandbox.
- Configuration changes onlyMemory, DATABASE_HOST, or readiness probe settings. No model-generated shell commands.
- Two remediation attemptsFailures end in an explicit terminal state with saved evidence.
- Honest execution labelsThe UI and RCA distinguish simulated validation from real Kubernetes verification.