// homelab

Isolation follows data sensitivity

15+ LXCs and VMs on my own hardware, a public dedicated server, an edge VPS — three tiers, one principle: convenience doesn't decide where a workload runs, the sensitivity of its data does. This setup isn't a toy next to the work; it's the infrastructure the work runs on.

// not a showroom, production

Three tiers

fig. 02tier 1 · on-premsensitive databehind NATno ports open into the housetier 2 · public dediclient volumemulti-tenantthe only public HTTP edgetier 3 · edge vpsstatic egress IPkeys onlyrevocable keys, no dataencrypted tunnel · outbound onlyisolation follows data sensitivity

deliberately abstracted :: no hostnames, no IPs, no topology — the principle is public, the inventory is not

// categories, not inventory

What runs here

Storage & backups

ZFS mirror, daily snapshots with verify, encrypted off-host backups — the NAS only ever sees ciphertext.

DNS & network

own resolver with adblock, secondary resolver, everything internal with real certificates.

Document pipeline

two Paperless instances (private/business) plus my own RAG search across the documents.

Agents

several self-hosted agent instances, skills as versioned infrastructure, LLM out of the action path.

Media & household

media server, photos, password manager, recipes — the household's cloud replacement.

Monitoring

metrics, uptime, dashboards — measurement before optimization.

// pinned to the wall

Principles

  • Isolation follows data sensitivity — sensitive data never leaves the house.
  • Never unencrypted onto the NAS: storage only sees ciphertext.
  • Measure, then optimize — measure first, touch second.
  • Backups are proven, not assumed: snapshot before every intervention.
  • LLM out of the action path: models produce text, deterministic code acts.