NVIDIA DGX Spark + Cloud Nemotron

Self-healing
Kubernetes copilot

An edge AI agent that detects cluster anomalies on-device and reasons over your entire infrastructure in the cloud — then fixes it before you wake up.

<3s
Detection
87%
Auto-resolved
0
False escalations
How it works
Edge detect. Cloud reason. Auto-fix.
Raw telemetry stays on-device. Only structured incident payloads reach the cloud for topology-aware remediation planning.
Edge · DGX Spark
W
Event watcher
Streams kubectl events, pod logs, and resource metrics continuously from the node
N
Nemotron edge
Classifies anomalies locally — CrashLoopBackOff, OOM kills, scheduling failures, resource pressure
P
Incident packager
Structures findings into a typed payload with severity, affected resources, and event timeline
→
Cloud · Nemotron
T
Topology reasoner
Cross-references the incident against Terraform state, node pool config, PDBs, and cluster dependencies
R
Remediation planner
Generates a ranked action plan — scale pool, adjust PDB, taint/drain, rollback deployment
→
Action layer
⚡
Auto-apply
Safe fixes execute immediately — pod restart, HPA adjustment, node cordon
!
Escalate to Slack
Risky actions surface as Slack threads with diff preview and one-click approval
Capabilities
Built for platform teams
◉
Continuous watch
Sidecar agent parses every event, log line, and metric on the node. No polling lag, no missed signals.
◈
Terraform-aware
Reads your IaC state to understand what should exist vs what does. No config drift surprises.
◇
Safe-by-default
Actions are tiered: safe ops auto-execute, destructive ops require human approval via Slack.
◎
Sub-3s detection
Edge inference catches anomalies before the next Prometheus scrape interval fires.
◉
Data stays local
Raw logs and metrics never leave the node. Only structured incident summaries reach the cloud.
◈
Incident memory
Learns from past incidents to reduce false positives and refine remediation strategies over time.
Live demo
Watch it heal
Force-delete a pod on stage. Watch the terminal cascade from red to green in under 3 seconds.
infraagent — prod-cluster-01
$ kubectl delete pod payment-svc-7f8b9 --force
■ [edge] CrashLoopBackOff detected → payment-svc-7f8b9 (ns: prod)
■ [edge] Packaging incident: severity=high, restarts=5, oom=false
■ [cloud] Correlating with terraform state... 3 replicas expected
■ [cloud] PDB allows 1 unavailable. Current: 1 unavailable. Safe to act.
■ [action] Auto-applying: kubectl rollout restart deploy/payment-svc
■ [action] ✓ Deployment healthy — 3/3 pods running (resolved in 2.4s)
NVIDIA DGX Spark Nemotron GKE Terraform Kubernetes API Slack API Prometheus Python