A record of what I've actually built

My homelab runs a dozen self-hosted services: files, media, network filtering, a ticketing site, several websites. All of that needs constant, careful attention: monitoring, updates, backups, incident response. I'm one person, with school and an apprenticeship on top, so I can't watch it around the clock, and paying for managed IT support wasn't an option.

HQ: a small "company" of specialized AI colleagues (built on Claude), each responsible for one slice of the infrastructure, network, storage, security, websites, automation, and so on, the same way a real ops team would split the work. They coordinate through a shared ticket board, so problems get picked up, worked and closed instead of getting lost in one long chat log.

The part I'm most proud of is the guardrails, not the AI itself: a gatekeeper program is the only way any colleague can reach the actual servers, it checks who is allowed to do what, and it takes an automatic backup before anything risky. Planned changes only happen in a quiet overnight window unless something is actually broken, and anything big enough to matter needs my explicit yes or no from my phone before it happens.

Cloud hosting for personal projects, files and media adds up in cost over time, and it means handing my data to someone else's servers. I wanted one place to run everything I build and use, self-hosted, properly backed up, and reachable from outside the house without opening ports on my home router (a real security risk if done carelessly).

Three physical machines clustered together with Proxmox, each one hosting several virtual machines and containers. Every guest gets a predictable IP address based on its own ID, so nothing needs a lookup table to find. One machine is dedicated to backups: it takes a snapshot of every guest every night and keeps a rolling history, so a broken update or a mistake is always recoverable.

Instead of opening ports on my router, every service that needs to be public goes out through its own encrypted tunnel to Cloudflare, so nothing on my home network is directly reachable from the internet. The HQ system described in the entry above runs on top of this and handles most of the day-to-day upkeep.

Most student project briefs invent a fake client with fake requirements. I wanted practice building for a real, specific need instead: a local business that actually needed working software, and a personal idea I was curious to build end to end myself.

Warenerfassung ("goods capture") for an Asian grocery store in Schaffhausen: staff scan a barcode and the app looks the product up, enriches it from Open Food Facts when the store doesn't have the data yet, and tracks how complete each product's record is so gaps don't get lost. It also ships as a desktop app (Electron), since a shop counter isn't always a great place for a browser tab.

Connection Room is a smaller, personal project: a real-time room where everyone who joins can see how they're connected to each other as a live graph, built to see how far I could push WebSockets for something that has to feel instant, not just "eventually consistent."

A lot of day-to-day admin, mine and small clients', is the same small set of actions repeated over email: read something, decide what it means, do something about it. Doing that by hand doesn't scale and is easy to let slip.

Mail Agent: a Python service that watches a mailbox and hands each message to a small set of specialised agents, split cleanly into layers (core logic, integrations with outside services, and the agents themselves), the same "split the work into specialists" idea behind HQ, just smaller and for one job instead of a whole homelab. It runs containerised so it's easy to move between machines without reinstalling anything by hand.