<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Infrastructure · ArchWorks</title><link>https://archworks.co/categories/infrastructure/</link><description/><language>en</language><lastBuildDate>Sun, 14 Jun 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://archworks.co/categories/infrastructure/index.xml" rel="self" type="application/rss+xml"/><item><title>A knowledge graph built from everything I've typed</title><link>https://archworks.co/posts/knowledge-graph-from-my-own-words/</link><pubDate>Sun, 14 Jun 2026 00:00:00 +0000</pubDate><guid isPermaLink="true">https://archworks.co/posts/knowledge-graph-from-my-own-words/</guid><description>RAG over your notes gives an AI a pile of text. A knowledge graph gives it entities and relationships. Here is how I build one only from my own words: collect, extract triples, dedupe, normalise, and read and write it live during a conversation.</description></item><item><title>Self-hosted LLMs and the context discipline that makes them work</title><link>https://archworks.co/posts/self-hosted-llms-subagents-skills/</link><pubDate>Sun, 07 Jun 2026 00:00:00 +0000</pubDate><guid isPermaLink="true">https://archworks.co/posts/self-hosted-llms-subagents-skills/</guid><description>Five layers on top of a local model: a single-slot model swap, model tiering, narrow agents with isolated context, skills that wrap an API as one tool call, and a memory layer that learns from its own runs. The discipline matters more than the model size.</description></item><item><title>Banking apps on stock LineageOS + microG in 2026</title><link>https://archworks.co/posts/microg-banking-apps-2026/</link><pubDate>Tue, 19 May 2026 00:00:00 +0000</pubDate><guid isPermaLink="true">https://archworks.co/posts/microg-banking-apps-2026/</guid><description>DEVICE-grade Play Integrity through microG works if you flip one toggle that defaults to block. The keybox is a weekly rotation, and every banking app layers its own checks on top. Here is the minimum stack that holds and what I tested as useless.</description></item><item><title>Handing my only GPU to a VM, and getting it back</title><link>https://archworks.co/posts/handing-my-gpu-to-a-vm/</link><pubDate>Tue, 28 Apr 2026 00:00:00 +0000</pubDate><guid isPermaLink="true">https://archworks.co/posts/handing-my-gpu-to-a-vm/</guid><description>Moving your only GPU into a Windows VM and back works. It also nukes your Plasma session every single time. I ran it as an experiment for months - here is the real cost and the layout that sidesteps the whole problem.</description></item><item><title>Ansible-Bootstrap: declarative Linux hosts across any distro and hypervisor</title><link>https://archworks.co/posts/ansible-clean-bootstrap/</link><pubDate>Wed, 22 Apr 2026 00:00:00 +0000</pubDate><guid isPermaLink="true">https://archworks.co/posts/ansible-clean-bootstrap/</guid><description>Describe a host as one block of YAML, run one playbook, get a configured server - and the same playbook does it to a whole fleet, the same way every time. The inventory says what a host should be; the AWX job log says what was done to it, when, and by whom.</description></item><item><title>Three-layer GitOps on K3s, in production</title><link>https://archworks.co/posts/kubernetes-gitops-layers/</link><pubDate>Sat, 18 Apr 2026 00:00:00 +0000</pubDate><guid isPermaLink="true">https://archworks.co/posts/kubernetes-gitops-layers/</guid><description>A 6-node K3s cluster as one git repo. Push to master, ArgoCD picks it up, nothing gets deployed by hand. Three layers, sync waves, and the parts that took longest to get right.</description></item></channel></rss>