ROUTINE / HIGH-VOLUME → EFFICIENT TIER
Triage, health sweeps, maintenance checks, cleanup, notifications, and other scheduled workflows use efficient models with scoped permissions.
AI HOMELAB — ENGINEERING DEEP DIVE
One Ubuntu host runs the core services. Two gateways coordinate five purpose-built agents with separate roles, model tiers, and permissions. Family and friends use the services; the operational complexity stays behind the interface.
Snapshot updated
The Operations Relay handles routing and schedules; the Inference Relay handles stronger reasoning tasks. Each agent receives only the tools and permissions its role requires.
▸ CLICK AN AGENT TO OPEN ITS DOSSIER
Routine, high-volume automation uses efficient open-weight models. Frontier-class models are reserved for complex engineering, security-sensitive work, and analysis that benefits from deeper reasoning.
Triage, health sweeps, maintenance checks, cleanup, notifications, and other scheduled workflows use efficient models with scoped permissions.
Infrastructure changes, security-sensitive operations, complex coding, and research analysis are routed to stronger models and require higher-trust tools.
Public-facing access travels through Cloudflare Tunnel. Administrative access uses restricted SSO or an encrypted Tailscale mesh.
Cloudflare Tunnel provides outbound-only access with WAF and TLS, without exposing router ports.
Cloudflare Access with a restricted Google SSO allow-list, or an encrypted Tailscale mesh.
Zero forwarded router ports, fail2ban active, and sensitive files restricted through host permissions.
The pipeline coordinates request intake, search, download, subtitles, import, and playback without requiring each user to understand the services underneath.
Request
Jellyseerr · family portal
Match + grab
Sonarr / Radarr
Search
Prowlarr · 6 indexers
Download
encrypted client
Subtitles
Bazarr
On screen
Jellyfin · refresh after import
Five agents have distinct responsibilities and permission boundaries, coordinated across two gateways.
Each agent remembers its own past work and nothing of anyone else's. A lookup runs two searches at once - one on meaning, one on exact wording - so a question finds the right note whether it repeats the original phrasing or not. Around 799 notes are indexed, and the embeddings are computed on the host, so operational memory never leaves the network. Built on sqlite-vec and SQLite FTS5 with a local EmbeddingGemma model.
Efficient open-weight models handle routine automation; frontier-class models are reserved for engineering and analysis that need stronger reasoning.
No forwarded router ports. Access uses outbound-only tunnels, restricted SSO, or an encrypted mesh.
Uptime Kuma runs 31 active health checks while the watchdog summarizes abnormal results and routes them for review. Health sweeps, maintenance checks, and research summaries run on defined schedules with scoped permissions.
Nightly restic snapshots use 7-day, 4-week, and 6-month retention with scheduled integrity checks.