Lutflow Sentinel

Sentinel is a lightweight, read-only agent that runs in your Kubernetes cluster. It collects GPU and inference metrics from your workloads for cost and efficiency analysis.

Current status: assisted Beta

Sentinel's hosted dashboard is currently available through an assisted onboarding process with our team — it is not yet self-service. If you're interested in piloting Sentinel with your workloads, contact us and we'll walk you through setup together.

The public chart and image below are a BYOK (Bring Your Own Kafka) technical preview: advanced users can point Sentinel at their own Kafka-compatible cluster to inspect the telemetry it produces. This preview mode does not connect to a Lutflow dashboard.

What Sentinel Does

Scrapes GPU metrics (power, utilization, memory) via DCGM
Collects token throughput from vLLM, TGI, or BentoML
Emits telemetry as structured events (JSON/Avro)

What it does NOT do:

Does not read prompts, completions, or any payload content
Does not modify your workloads (read-only by default)
Does not require cluster admin privileges

Prerequisites

Kubernetes 1.25+
Helm 3.10+
NVIDIA GPU nodes with DCGM exporter
Inference workloads (vLLM, TGI, or BentoML) exposing /metrics

Try the BYOK preview

This installs Sentinel in local measurement-only mode — no external transport, metrics stay on the agent's own /metrics endpoint:

helm install lutflow-sentinel oci://ghcr.io/lutflow/charts/lutflow-sentinel \
  --namespace lutflow --create-namespace \
  --set global.tenantId="your-tenant-id" \
  --set kafka.enabled=false

If you want to export telemetry to your own Kafka-compatible cluster instead, we recommend referencing an existing Kubernetes Secret (kafka.existingSecret) rather than passing credentials via --set, since --set values are visible in shell history and Helm release metadata.

Label Your Workloads

Add the discovery label to pods you want to monitor:

apiVersion: v1
kind: Pod
metadata:
  labels:
    lutflow.io/enforced: "true"
    lutflow.io/model: "llama-3-8b"      # optional: model name
spec:
  containers:
    - name: vllm
      ports:
        - containerPort: 8000

Verify Installation

kubectl get daemonset -n lutflow
kubectl logs -n lutflow -l app.kubernetes.io/name=lutflow-sentinel --tail=20

Supply Chain Verification

Sentinel images are signed with Cosign. Verify before installing:

cosign verify ghcr.io/lutflow/sentinel:0.1.5 \
  --certificate-oidc-issuer=https://token.actions.githubusercontent.com \
  --certificate-identity-regexp="github.com/Lutflow/sentinel"

Uninstall

helm uninstall lutflow-sentinel -n lutflow
kubectl delete namespace lutflow

Questions?

Security architecture: Security & Trust
Support: support@lutflow.dev

Sentinel (Helm)