minipc-01 · fedora · 28 containers · 38 ansible roles
One box.
Everything on it.
Media, files, a local language model and the automations around it, all on one mini PC in a closet. The whole machine rebuilds from one Ansible playbook.
The board
Every part of the machine and what runs on it. Point at a part to see its services, or at a service to see where it lives.
- language model, 32B parameters, ~3.5B read per word
- embedding and reranking models for document search, 3.5 GB
- the other 25 containers and Fedora together
The model shelf
One model, always in memory. The others that were tried here have been deleted. Bar length is writing speed, measured on this box.
- Clover 35B Always loaded. Every automation, document search and the coding agent. A diffusion model that can write in blocks; here it still writes word by word, until llama.cpp learns its fast mode. 7.2 tok/s
Services
Sorted by how far each one is from the internet. The colour is the distance. Apps marked built here were written for this box and deploy themselves on a push.
publicon the open internet, behind TLS
home networkreachable from the LAN and the VPN
-
Grafana + Prometheus
Dashboards for CPU and GPU load, both drives' temperature and fill, and the model's live reading and writing speed.
-
Forgejo
Git hosting for every project on the box, with CI. Each push is built, tested and scanned in a sandboxed Docker, and an app that passes deploys itself, and rolls itself back if it comes up broken.
-
n8n
The automation engine, wired to the local model. Telegram reaches it through a single public webhook path.
-
Doc search
Search over the box's own notes and code. Ask why a decision was made and get the answer with the file and line it came from. A second small model re-reads the best candidates before they are returned.
-
Car service book built here
Service history, receipts and photos for the car, with reminders by kilometres or months, whichever comes first.
-
Networking course built here
An interactive CCNA course written for a class: 57 lessons, a router and switch simulator with 61 labs, flashcards and a timed practice exam. Installs on a phone and keeps working offline.
-
Production planner built here
A planning demo for a food producer: what can be delivered against what was ordered, recomputed as the day changes. The local model explains each plan in plain words.
-
Plex
The second media server, for the TVs in the house.
-
NFS
The media drive shared with every machine on the network.
-
Pi-hole
Ad and tracker blocking for every device, at the DNS level.
internalprivate Docker networks, never published
-
llama.cpp
Clover 35B on the Vega 8 through Vulkan, 64k context. A mixture of experts: all its parameters stay in memory and about 3 billion are read per word. It is the only model on the box. Prompts never leave it.
-
Embedding + reranking models
Two small models for search: one turns text into vectors, the other scores each candidate against the question. Kept apart from the main model so tuning them never disturbs it.
-
Qdrant
Meaning and exact keywords indexed side by side, so a plain question and a variable name both find the right paragraph.
-
Transmission
Downloads, driven entirely from Telegram. No web interface at all.
-
MariaDB
The database behind Nextcloud, the planner and the car service book, on a network nothing else can see. Forgejo keeps its own PostgreSQL.
-
SearXNG
Web search for the automations and the coding agent, returning short extracts instead of whole pages.
Automations
n8n workflows call the local model. Everything is worked out on the box; only the finished message leaves it, to Telegram.
-
daily 07:30
Morning briefing
Weather, today's calendar, headlines and how the server slept, written up by the model as one message.
-
daily 07:30
Car reminders
Services, inspections and registration dates coming due, from the service book. Silent when nothing is.
-
mondays 08:00
Image update watcher
Checks every pinned container image for a newer release; the model writes the change notice.
-
mondays 08:15
Security report
The week's SSH attempts and fail2ban bans, summarised.
-
nightly 04:00
Doc index canary
Re-indexes the documentation and checks it is still there. Silent when healthy.
-
on request
Media sorter
Send a magnet link. When the download finishes the model names it, splits season packs by episode and files it in the library.
Build log
Everything that changed, newest first.
-
2026-11-16
One model left
The shelf was cleared. Clover 35B is now the only language model on the box and the one everything uses; Nemotron, Ornith, Laguna, LFM and the dense Qwen were deleted, which gave the media disk back about 81 GB. Jellyfin can now save the subtitles it downloads next to each film, decodes two more video formats on the graphics chip, draws preview thumbnails on the seek bar and reports its numbers to the monitoring.
-
2026-11-13
Nine quiet days, found
A routine check of the whole box found the language model had been off for nine days. It was stopped by hand before a Docker upgrade on 3 November, as the notes require, and the reboot that followed left it stopped, because a container stopped on purpose stays stopped. Nothing complained, since every web page was still up. The model now starts on every boot whatever state it was left in. The same check found the weekly dependency updater finishing green without looking at a single repository since late September, over a search pattern its engine could not read. Both are fixed, ten containers were brought up to date the same afternoon, and this page got a model shelf, a light theme and a log that reads newest first.
-
2026-11-12
A course with its own routers
A complete networking course for a class preparing for the CCNA exam, written and served from the box: 57 lessons across all six exam domains, 395 questions, flashcards on a spaced schedule and a timed practice exam weighted like the real one. Its centrepiece is a simulator of routers and switches that runs in the browser, with 61 labs checked against the state of the devices rather than the commands typed, and every lab is solved automatically in CI so a broken one cannot ship. It installs on a phone, works with no connection and sends the answers back when one returns.
-
2026-11-03
Everything one version newer
An evening of updates with a backup before each: Forgejo to a new major version, the automation engine, both media servers, the dashboards, the DNS filter and the search engine. llama.cpp moved to a newer build only after the same prompt was timed before and after, and the speed came out the same to within noise.
-
2026-10-30
A shelf of models
Two more models were measured on real server tasks and kept for when they suit. A small fast one answers quick lookups at 16 words a second but loses the thread on anything long, so it gets a short leash: no shell of its own and a hard limit on how many steps it may take. A middle one writes a little faster than the resident model and matches it on short tasks, then slows as the conversation grows. Neither replaced the resident model; both are one word away.
-
2026-10-24
The car gets a service book
A small app for the car's whole history: every service and receipt, photos, and intervals that count down by kilometres or months, whichever comes first, plus the legal dates. Anything coming due arrives with the morning briefing. Like everything else here it deploys on a push, with its database dumped before each new version goes in.
-
2026-10-13
A faster brain, sharper search
A day of bake-offs. Five candidate models were measured against Ornith on speed, quality and seven real agent tasks run on this box, and Nemotron 3.5 Lightning came out on top: about a fifth faster to answer, and almost no slower deep into a long conversation, because only six of its 52 layers have to look back over everything said so far. It is now the model that is always loaded; Ornith and a careful dense model wait on the shelf for when they are asked for. Document search gained a second model that re-reads the best dozen candidates before answering, which put the right note in the top three for ten of twelve test questions instead of eight. The coding agent got guard rails measured on the same tasks: it can no longer run a server command on the wrong machine, and it is told to stop digging and answer once it has the facts. The same evening the whole box was brought up to date: a new kernel, Docker, llama.cpp, MariaDB 13, PostgreSQL 18 and two major versions of Nextcloud, each one backed up first and checked after. The receipt-reading vision model and the browser chat front end were retired as unused, and the box's own Python services moved to Python 3.14.
-
2026-09-28
Push to deploy
Every project moves into its own repository on a Git server on the box, and a push now does the rest. The code is built and tested in a sandboxed Docker that cannot touch the rest of the machine, checked for known vulnerabilities, and started once against an empty database to prove it boots. Only then does the box swap it in, and if the new version does not answer like itself within a couple of minutes, the old one is put straight back. On the first run the vulnerability check caught a web framework 62 security fixes behind; it was updated the same hour. Dependency updates now arrive as a weekly pull request that has to pass the same checks.
-
2026-09-27
A second drive
An NVMe drive goes into the empty M.2 slot and the whole media library moves onto it — films, shows, music and the download folder together, because the sorter files a finished download by linking it rather than copying it, and that only works when both live on the same drive. The system drive gets back more than half of its space, and the models, databases and everything else stop competing with the library for it. Both drives now report their own temperature and fill level on the dashboards and in the morning briefing, with separate limits, because an NVMe drive runs warmer by nature.
-
2026-08-23
One model, newer
The brain moves to the next version of the same model — same shape and same settings underneath, only retrained weights, so nothing else on the box had to change. Everything else is cleared off the disk with it: the older versions kept as a way back are gone, and what remains is this model and the document reader, nothing more. Two honest caveats, because the notes say so too. No side-by-side speed or quality test was run first, so "newer is better" is a reasonable expectation rather than something measured here. And with no spare model on disk, going back means downloading one again instead of changing a single line.
-
2026-08-21
Making room
An accounting of where the memory actually goes: the model is 82% of it, and everything else on the box together is a fraction of one model. The browser chat window turned out to be the largest thing that was not the model — partly because it was quietly loading a second embedding model of its own, duplicating a job the retrieval stack already does better. That got repointed at the real embedder, and the chat front end retired. The same model still answers the coding agent, the automations and the document search; only the browser window is gone. Nothing is deleted — accounts and history sit on disk behind one flag.
-
2026-08-20
Four times the model, faster answers
A 35-billion-parameter model takes over from the 9B — and it is quicker on every measure, because it wakes under a tenth of itself for any one word. Prompt reading gains a third, writing speed rises by four fifths, and it scores measurably better on the standard quality test rather than trading accuracy for speed. The cost is space: three times as much memory held permanently, which meant lifting the ceiling the graphics chip is allowed to borrow from system RAM and taking a reboot to do it. The 9B stays on disk as the way back, unused unless something asks for it by name.
-
2026-08-20
The box reads its own notes
Everything written about this machine — every Ansible role, every design note, every benchmark — becomes searchable by meaning as well as by name. A second small model, permanently in memory, turns the text into vectors; a store keeps those beside a keyword index and merges both rankings, so "why is the cache unquantized" and an exact variable name each land on the right paragraph. Answers come back with the file and line they came from. The coding agent on the laptop now asks the box about the box.
-
2026-08-14
Ornith comes back
The 9B model that ran the workflows back in July returns as the one brain for everything — chat, the coding agent and all five automations. Same size and the same speed as the model it replaces, because underneath they are the same architecture differently tuned; the swap is a bet on how it writes, not on how fast it runs. Roughly eight gigabytes resident, which leaves the long-conversation cache plenty of room. The previous model stays on disk, one line away, in case the new one disappoints.
-
2026-08-13
The iGPU wakes up
Ollama gives way to llama.cpp, and inference moves off the CPU onto the Vega 8 through Vulkan — prompt processing gets five and a half times faster. One model now serves everything, held in memory permanently so it remembers the start of a long conversation instead of re-reading it every turn. Chat, the document reader and every automation share it.
-
2026-07-22
New brain
Ornith 35B takes over from Qwen3.6 as the workflow and chat model. Same tokens per second, but it loads in three quarters of the time and answers straight away instead of thinking out loud first — so replies arrive sooner without the box working any harder.
-
2026-07-18
Chat with the box
Open WebUI puts a face on the local LLM — chat from any browser, anywhere, behind Cloudflare Access. Tuned for CPU inference: a "hello" that first took five minutes now answers in three seconds.
-
2026-07-18
Memory doubled, models tripled
A second RAM stick unlocks dual-channel bandwidth — inference speed nearly doubles overnight. A three-way bake-off crowns Qwen3.6 35B as the new workflow brain, and the lift vision model steps up to a near-lossless quant with double the context.
-
2026-07-11
Sharper brain, faster pipeline
All LLM workflows switch to Ornith 9B after a CPU benchmark bake-off. The media sorter goes event-driven — Transmission fires a webhook the second a download completes — and learns to sort whole season packs, episode by episode.
-
2026-07-09
Morning briefing
The daily digest grows up: weather, Google Calendar events and news headlines join the server metrics, composed by the local LLM into one 07:30 Telegram message.
-
2026-07-07
Media pipeline
Transmission wired to Telegram: send a magnet, the LLM classifies the release and files it into the Plex/Jellyfin library. Downloads, sorting and naming, all on-device.
-
2026-07-07
Document AI
Datalab lift 9B vision model self-hosted on CPU. A gateway rasterizes PDFs, extracts schema-constrained JSON and renders the result back to PDF — driven from Telegram.
-
2026-07-06
AI era
Ollama serving Qwen3 locally, n8n automation engine, and three LLM-powered workflows reporting to Telegram.
-
2026-07-05
Eyes and cloud
This website. Uptime Kuma, hardware sensors, Prometheus + Grafana dashboards. Nextcloud on MariaDB with Redis.
-
2026-07-04
Media era
Jellyfin with AMD VAAPI hardware transcoding, Plex, NFS media share — plus nginx reverse proxy and Let's Encrypt DNS-01 certificates.
-
2026-06-04
git init
Repo born. Base roles: users, SSH hardening, firewalld, Docker, Pi-hole, fail2ban, automatic updates.
Stack
- os
- Fedora Server · SELinux enforcing
- containers
- Docker + Compose · 28 running
- automation
- Ansible · 38 idempotent roles · secrets in Ansible Vault
- ci/cd
- Forgejo + Actions in Docker-in-Docker · tests, audit, image scan · Renovate · pull-based deploys with health-gated rollback
- inference
- llama.cpp + llama-swap · Nemotron 3.5 Lightning 30B hybrid Mamba mixture of experts, 128k context · five more models on demand
- retrieval
- Qwen3 embeddings + reranker · Qdrant hybrid dense + sparse · key-gated FastAPI gateway
- apps
- FastAPI · React + TypeScript · SQLite or MariaDB · installable, offline-capable where it helps
- ingress
- nginx · Let's Encrypt DNS-01 via Cloudflare
- data
- MariaDB 13 · PostgreSQL 18 · Redis 8 · isolated Docker networks
- security
- SSH hardening · firewalld · fail2ban · automatic updates
- monitoring
- Prometheus + Grafana · smartd on both drives · Telegram alerts