AI grounded in your live cluster state. Approval-first on every mutating operation.
CyPilot is a conversational AI assistant for operations and troubleshooting. Powered by an on-prem LLM (phi3:medium-128k) + RAG over your knowledge base + real-time cluster state. Intent parser classifies risk (READ_ONLY / LOW_RISK / DESTRUCTIVE). LOW_RISK operations auto-execute; DESTRUCTIVE always requires explicit approval.
'VM X is unreachable' → CyPilot runs diagnostics, suggests fix, runs LOW_RISK remediation if confidence > 0.85. Operator approves anything risky.
Junior SREs ask CyPilot how to do things. Grounded in your actual runbooks (RAG-indexed). No more 'wait for the senior to come back'.
'How are PHI records encrypted in this cluster?' → CyPilot answers with cluster-specific config + audit trail, not generic docs.
During incidents, CyPilot pulls correlated signals from CyOptics + Prometheus + logs and proposes the most likely root cause. With citations.
Tenant admins ask CyPilot 'show me my running VMs' or 'create a new dev environment with 4 vCPUs'. LOW_RISK approvals route to tenant policy.
Knowledge base searches return facts from your live state. Documentation drafts grounded in actual configurations.
phi3:medium-128k. No prompt data leaves your cluster. Sovereign AI.
90+ default knowledge entries; ingests your runbooks. nomic-embed-text embeddings, cosine similarity retrieval.
DESTRUCTIVE actions never auto-execute. LOW_RISK allowlist is narrow + explicit.
Risk classification on every request. CREATE/DELETE/ACTION/QUERY/TROUBLESHOOT/WORKFLOW intent shapes.
Every action audit-logged with actor, intent, parameters, outcome. Reversible where possible.
Approvals route per tenant policy. DCOS_ADMIN vs TENANT_ADMIN vs USER scopes.
30-day proof of concept on two idle servers. We bring the SE. You bring the use case.