Initializing Gaussian Splat Engine...
SYS_LOC: K8S.EDGEPRIME.IO
MODULE: HERO
FPS: 60.0
RENDER_METHOD: 3D_GAUSSIAN_SPLAT_PHOTOGRAMMETRY
STATUS: ONLINE
CLUSTER DE PRODUCTION EN DIRECT · 8 NŒUDS

Quittez le Cloud.
Maîtrisez votre infrastructure.

Welcome to the digital mind of a Cloud-Native Architect. Scroll through my IT experiences to explore how I build high-availability Kubernetes platforms, secure zero-trust workloads, and scale edge AI clusters.

kubectl get nodes
CLUSTER: *.apps.edgeprime.io
NAME ROLE RAM GPU OS / ARCH
controlplane-100 control-plane 16 GB None Talos v1.12.6 (amd64)
controlplane-101 control-plane 8 GB None Talos v1.12.6 (amd64)
controlplane-102 control-plane 16 GB None Talos v1.12.6 (amd64)
worker-103 worker 24 GB None Talos v1.12.6 (amd64)
worker-104 worker 24 GB None Talos v1.12.6 (amd64)
gpu-105 worker 96 GB RTX 2060 Talos v1.12.6 (amd64)
worker-106 worker 16 GB None Talos v1.12.6 (amd64)
gx10 worker 128 GB GB10 Blackwell Ubuntu 24.04 (DGX OS) (arm64)
API SERVER: https://k8s.edgeprime.io:6443
ALL NODES READY
🏗️
TALOS LINUX · CILIUM eBPF · K8S v1.35.3

Core Platform & Network CNI

A minimal, production-grade bare-metal architecture built with an API-managed immutable OS and kernel-level networking.

CILIUM eBPF INGRESS FLOW
[INTERNET] --(HTTPS:443)--> [CILIUM GATEWAY API]
└─ Terminate TLS (Let's Encrypt + Cloudflare API)
[CILIUM GATEWAY] --(eBPF Direct Route)--> [HTTPRoute: *.apps.edgeprime.io]
└─ Forwarding downstream to ClusterIP Service (Port 80/443)
[CILIUM CNI] Connection Established -- Path resolved in 0.12ms

Talos Linux

Immutable, security-hardened Linux distribution. No SSH, no systemd, managed entirely through gRPC.

Cilium CNI

eBPF-based container networking, replacing kube-proxy with kernel-level routing and Gateway API ingress.

Hubble

eBPF network flow observability, mapping internal microservice dependency streams in real-time.

🔄
ARGOCD · TERRAFORM · RENOVATE

GitOps & Infrastructure IaC

Fully declarative GitOps control plane where every system configuration, secret sync, and network policy is versioned.

GIT RECONCILIATION LOGS
[2026-06-22T21:12:05Z] INFO ArgoCD controller polling repository hkboujrida/PKC
[2026-06-22T21:12:08Z] DIFF Out-of-Sync: Helm values file edited in branch main
[2026-06-22T21:12:11Z] SYNC Reconciling resources ... Apply SecretSync (eso-vault-sync)
✓ STATE SYNC COMPLETED · 0 DIFFERENCES

ArgoCD App-of-Apps

Continuous delivery engine managing 45+ applications across 35+ namespaces with automated reconciliation.

Terraform Modules

Infrastructure as code managing Vault credentials, Keycloak client integrations, and Grafana data sources.

Renovate Engine

Automated dependency manager tracking and validating Helm chart bumps and Docker tag updates.

🛡️
KEYCLOAK · VAULT · KYVERNO · KUBESCAPE

Zero-Trust Security & Policies

Comprehensive cluster defense-in-depth, incorporating automated identity, secrets injection, and runtime audits.

HASHICORP VAULT STATUS
Raft Cluster Size: 3 nodes
HA Mode: Active Leader
Seal Status: Unsealed
Auto-Unseal: webhook-event
ESO Secret Bridges: 20 Sync Paths
Compliance: CIS/NSA Audit 98.4%

Keycloak IAM

Centralized Single Sign-On (SSO) with custom theme, protecting all admin portals via OIDC/SAML.

HashiCorp Vault HA

High-availability Raft cluster serving as the encrypted database of truth for API keys and certificates.

Kyverno Policy Engine

Kubernetes admission controller enforcing security compliance, mutating resources, and blocking root containers.

📊
MIMIR · LOKI · TEMPO · PYROSCOPE · ALLOY

Grafana LGTM+ Telemetry Stack

Unified telemetry pipeline gathering all five signals (metrics, logs, traces, profiles, and costs) across the nodes.

PROMETHEUS LIVE METRICS
QUERY: sum(rate(container_cpu_usage_seconds_total[5m])) by (namespace)
ns: kube-system (Cilium, CoreDNS) 12.4% CPU
ns: entertainment (vLLM, Immich) 64.8% CPU
ns: observability (Mimir, Loki) 22.1% CPU

Grafana Mimir & Loki

Scalable metrics and log engines providing long-term retention and fast multi-tenant querying.

Tempo & Pyroscope

Correlated distributed tracing and continuous CPU/Memory profiling to diagnose microservice latency.

HolmesGPT AI Operator

Autonomous AI-powered troubleshooting assistant investigating cluster anomalies using Minimax LLM.

💾
LONGHORN · CLOUDNATIVEPG · DRAGONFLY · GARAGE S3

Storage & Distributed Databases

Resilient stateful storage layer with replicated block pools, HA database clusters, and secure offsite object storage.

LONGHORN 3-WAY REPLICA MAP
Replica-1
k8s-03 (NVMe)
● RUNNING
Replica-2
k8s-04 (NVMe)
● RUNNING
Replica-3
k8s-05 (NVMe)
● RUNNING

Longhorn Storage

Cloud-native block storage with active 3-way synchronous replica distribution across physical NVMe disks.

CloudNativePG

Enterprise-grade PostgreSQL operator handling automated replication, backups, and point-in-time recovery.

Garage S3 Object Storage

Lightweight distributed S3 object store providing secure backup targets for PG databases and Velero.

🤖
DGX BLACKWELL · AIBRIX · vLLM · NGC

Bare-Metal AI & GPU Workloads

Edge-AI computing cluster deploying local LLM inference models on bare metal with dedicated GPU hardware.

NVIDIA GB10 BLACKWELL TELEMETRY
Model serving: Qwen2.5-72B
Inference rate: 82.4 tokens/s
Cache Strategy: Prefix Cache
GPU Temp: 62°C
Power Usage: 380W / 700W
VRAM allocated: 72.4 GB / 128 GB

NVIDIA DGX Spark Node

Worker node powered by Grace Blackwell GB10 Superchip, featuring 128GB unified memory and ConnectX-7 link.

AIBrix Gateway

Advanced model serving control plane providing prefix-cache-aware routing and LLM request load balancing.

vLLM Inference Runtime

High-throughput LLM engines serving Qwen-2.5 and Llama models with PagedAttention optimizations.

Take Back Your Infrastructure

Explore the fully functional platform code, Kubernetes manifests, and automation scripts powering the PKC cluster.