[01]
AI compute
Alpha-5 runs private inference and the surrounding chat, retrieval, memory, search, and tool-serving stack on dedicated GPU hardware.
RTX 5090 · vLLM · LibreChat · Open WebUI · MCP
Compute · storage · virtualization · orchestration
The Hive.Operate itto learn it.
A self-hosted platform spanning GPU inference, virtualization, 100+ TB of ZFS storage, and a nine-node Kubernetes cluster—built to run AI and cloud-native systems on infrastructure I own end to end.
// platform status
●Git reconciled
●Local models serving
●Workloads observable
●ZFS pool online
32
GB GPU memory
102
TB ZFS pool
09
k3s nodes
10
CNCF certs
01 — Why it exists
Fluency through operations.
Cloud-native systems only become intuitive when you live with their failure modes. The goal was to build a durable environment where networking, storage, delivery, identity, observability, and application operations could be learned as one connected system. It is a laboratory, but it is operated with production habits: declarative delivery, observable behavior, repeatable recovery, and documented decisions.
02 — The whole platform
[02]
Virtualize
Proxmox provides the durable substrate for core services, virtual lab machines, automation targets, and selected cluster capacity.
Proxmox VE · QEMU · LXC · ZFS
[03]
Store
TrueNAS centralizes durable data in a 102 TB ZFS pool and serves application data, shared files, photos, and backups.
TrueNAS · ZFS · NFS · SMB
[04]
Orchestrate
A mixed nine-node Linux fleet runs k3s and containerd, turning the hardware estate into a programmable application platform.
k3s · containerd · Ubuntu · Debian
[05]
Deliver
Git is the control plane. Argo CD continuously reconciles platform services and product workloads into the cluster.
Argo CD · Gitea · AWX · Jenkins
[06]
Operate
Network flows, metrics, logs, traces, secrets, and certificates are visible and managed as parts of one system.
Cilium · Prometheus · Loki · Tempo · Vault
03 — System map
One estate. Distinct jobs. Deliberate boundaries.
Access + edge
Tailscale
Caddy + Traefik
Authentik
AdGuard
Foundation
ProxmoxVMs · LXC · lab environments
TrueNAS102 TB ZFS · NFS · durable data
Alpha-5 / AI
RTX 5090 inference
vLLM · LibreChat · Open WebUI · RAG · vector search · MCP
The Hive / k3s
Nine-node platform
GitOps · products · databases · observability · developer tooling
identity + routingcompute + stateworkload platforms
04 — Alpha-5 / Local AI
The machine at the center.
Alpha-5 is where the homelab stops being an infrastructure exercise and becomes an AI platform. A 24-core Intel Core Ultra 9 285K, 64 GB of system memory, fast local NVMe, and an RTX 5090 with 32 GB of VRAM run private model inference and the services around it.
vLLM serves the models. LibreChat and Open WebUI provide human interfaces. RAG, pgvector, MongoDB, Meilisearch, and MCP services supply retrieval, memory, search, and tools. It is a complete local AI workbench—not a GPU waiting for a prompt.
alpha-5 / hardware profile
- CPU
- Core Ultra 9 285K / 24 cores
- GPU
- GeForce RTX 5090 / 32 GB
- RAM
- 64 GB
- LOCAL
- 4 TB NVMe
- RUNTIME
- Ubuntu · Docker · vLLM
05 — What runs here
A platform for platforms.
The estate hosts the tooling needed to build and operate software, plus real product workloads: local AI, a multi-service practitioner platform, internal MCP services, developer environments, databases, home automation, identity, storage, and observability.
- ▹Alpha-5 local AI stack: RTX 5090 inference with vLLM, Open WebUI, LibreChat, RAG, vector search, and MCP tooling
- ▹Proxmox virtualization for Kubernetes, Home Assistant, identity, DNS, reverse proxying, data services, and isolated lab environments
- ▹TrueNAS with a 102 TB ZFS pool serving shared storage, photos, backups, and application data
- ▹GitOps delivery for 25+ applications with Argo CD
- ▹Distributed networking and service exposure with Cilium, Hubble, MetalLB, Traefik, and Kong
- ▹Persistent workloads on Longhorn and NFS-backed storage
- ▹Metrics, logs, and traces through Prometheus, Grafana, Loki, Tempo, and OpenTelemetry
- ▹Platform services including Vault, External Secrets, cert-manager, Gitea, AWX, Jenkins, and Coder
- ▹Ten CNCF certifications: KCNA, KCSA, PCA, OTCA, CGOA, CAPA, CCA, KCA, CBA, and CNPA
Constraints / Evidence
The conditions shaped the system.
- ◆Mixed commodity hardware with finite power, memory, storage, and GPU capacity
- ◆No managed cloud control plane: upgrades, recovery, networking, and state remain operator-owned
- ◆Stateful workloads span local disks, replicated volumes, NFS, and ZFS failure domains
- ◆Public evidence must prove operation without exposing topology or creating an access path
9
k3s nodes
A heterogeneous cluster operated as one application platform.
25+
GitOps applications
Desired state continuously reconciled through Argo CD.
102 TB
ZFS storage
Durable application data, files, photos, and backups.
32 GB
GPU memory
Private inference served locally from Alpha-5.
The result
The certifications validate the knowledge. The homelab proves I can design, integrate, and operate the whole system.
Talk infrastructure →