About.md

I build things to understand them. Sometimes they work. Sometimes they teach me something better than working would have.

I'm Christian T. Drieling - engineer, tinkerer, and collector of more side projects than I can name in one sitting. Based in Oldenburg, Germany.

By day I work with the kind of infrastructure people only notice when it breaks - virtual desktops, Microsoft cloud, identity, the parts of IT where downtime ends up in someone's calendar at 7am. By night I run a homelab where reliability is optional and learning is mandatory. The contrast keeps both honest.

This blog is where I write down what happened: the experiments, the infrastructure decisions, the AI rabbit holes I fell into at 11pm on a Tuesday. No polished theory - just applied curiosity. Try something, see what breaks, figure out why.

It started, slightly embarrassingly, because I kept opening new tabs to write things down anyway. At some point, creating this blog felt less ridiculous than another folder of experiment-final-FINAL-v3.md files.

What this is

The Context Window is a personal lab journal. I write about things I've actually tried, built, or broken - with enough detail that future-me (and maybe you) can reproduce or avoid it.

A few recent rabbit holes:

  • The memory got a door - the memory folder and the eleven tool servers my agents use lived on the machine the agent ran on, invisible from my phone, until a bad sync morning forced a rebuild: one self-hosted door in front of all of it, rules enforced by the server, reachable from the workstation and from the phone. The phone was the wish - the bad morning was the reason.
  • Markdown Mind - distilled three hand-built agent-memory systems into one shared kernel plus presets. The hard part wasn't the design; it was watching a small local model misread the setup until I stopped writing instructions and started handing it things to copy.
  • The memory folder - moved everything my agents know about me out of vendor memory features and into plain Markdown I own. The point was never the folder - it was that the model underneath became swappable.
  • A market of two - built a real, self-hosted household-finance tool for an audience of exactly two people. Not everything worth building needs a market.

Topics I keep coming back to:

  • AI that runs on hardware you own
  • The memory and tools your AI runs on, kept in plain files you own
  • Self-hosted infrastructure and the lessons it keeps teaching me
  • LLMs in practice - what works outside the demo videos
  • Automation, the home lab, and the small joy of it works on my machine

What this is not

  • A tutorial site. Things break mid-post and I'll tell you.
  • Thought leadership. I'm figuring this out as I go.
  • Up to date. I write when something is worth writing about.
  • Vendor-neutral. I use what works for me, not what's popular.
  • A substitute for the docs. Read those too.

How this is made

The projects, the decisions and the mistakes are mine - every post starts from something I actually built, ran or broke. The text is written with AI: agents draft a first version in my voice from my notes and logs, then I edit and approve every line before it goes out. Header images are generated from prompts I write, diagrams are drawn with AI help, and the software in these posts is largely built by AI coding agents under my direction - which is, more or less, what this blog is about. No ads, no sponsors, no affiliate links. Opinions are my own and not my employer's.

The stack

Local-first. Cloud where it earns its keep - and Claude has been earning it. This list ages fast - assume it has drifted by the time you're reading.

Hardware

MacBook Pro M4 Max - 128 GB unified memory
Mac Studio M4 Max - 64 GB, dedicated to local inference

Local inference LM Studio · oMLX · Ollama on Proxmox

Models I run locally

  • Text: Qwen3.8-27B · Qwen3.8 Flash Next · Gemma4 and friends
  • Voice (TTS): Qwen3-TTS · Voxtral-4B
  • Speech (STT): Parakeet-TDT
  • Embeddings & rerank: jina-embeddings-v5 + jina-reranker-v3

Automation n8n · Open WebUI Pipelines · Hermes Agent

Search & RAG Perplexica · Apache Tika · and the Jina stack

The agent layer

Local-first doesn't mean local-only. Claude Code is still my daily driver across the Mac fleet, including a couple of nodes that stay on 24/7 and just keep working while I sleep. The MCP agents I built myself now sit behind one self-hosted door - Denkarium, German for Pensieve - that the workstation and the phone reach through one connector, claude.ai on the phone included. The door enforces the house rules instead of hoping the client reads them. Behind it:

  • Calendar (CalDAV)
  • Mail (IMAP read, SMTP send)
  • Home Assistant
  • A self-hosted household-finance tool
  • A shopping list
  • Notifications to the phone

The memory layer is a git-backed Markdown tree now - Obsidian is still the editor, git is the transport, and every agent on every device works on the same repositories. Fitting, for a blog called The Context Window.

Claude (Anthropic) is the primary cloud model. OpenRouter covers the rest when I need a second opinion - GPT, Gemini, Mistral.

Uptime: mostly.

Get in touch

Code lives at github.com/cdrieling. Mail goes to blog@apanthos.com - corrections, ideas, or "this broke in an interesting way" are all welcome.