Field notes on systems, privacy, automation, and digital culture.https://facko.decortel.de/enhttps://facko.decortel.de/notes/routing-for-tokens/https://facko.decortel.de/notes/routing-for-tokens/I did not set out to learn LiteLLM, RAG, gateways, any of it. I was wasting tokens. Every question went out with the whole tool list attached, whether I asked about a virtual machine or something I could have grepped. For a while the compose file was just names I copied. The picture that stuck: chat talks to LiteLLM, LiteLLM talks to the model. RAG is not memory. It is a search over notes I already wrote, stuffed back into the prompt so the model can quote me instead of guessing. I have to load those files myself. They do not update on their own. I found that out by getting an old answer. The router came after I measured the waste. Keywords pick a smaller set of tools. Some two-part questions get split so the second half actually gets asked. It is string matching. I had been waiting for something cleverer, and that was the mistake. I did not need a smarter model to spend fewer tokens. I needed to stop sending the whole workshop when the question only needed one drawer.Sun, 13 Sep 2026 00:00:00 GMTLiteLLMRAGtokenshttps://facko.decortel.de/notes/gpu-passthrough-again/https://facko.decortel.de/notes/gpu-passthrough-again/The RX 7600 XT guest stopped responding again. The VM list still looked reassuring. A closer status check reported an internal error. This time the card was still on the PCI bus. Stopping the guest worked. Starting it with the GPU attached did not. The host reboot brought inference back. The failure happened while idle, which weakened the earlier theory about load being the trigger. Firmware polling and runtime power management became the next things to test. They are a working hypothesis, not a proven cause. “Running” is a process state. It is not a health check. A small inference request tells me more than the green status label.Thu, 03 Sep 2026 00:00:00 GMTProxmoxGPUrecoveryhttps://facko.decortel.de/notes/one-dashboard/https://facko.decortel.de/notes/one-dashboard/The Home Assistant sidebar had accumulated several ways to get to roughly the same controls. The useful change was making one room-oriented dashboard the starting point and moving the old ones out of the way. A missing sensor value also needed to stay missing. Turning “unavailable” into zero makes a tidy graph and a misleading room. The next pass is about physical behavior: the dimmer, the scene, and the dashboard should agree about what just happened.Wed, 02 Sep 2026 00:00:00 GMTHome Assistantinterfacesmaintenancehttps://facko.decortel.de/notes/quiet-systems/https://facko.decortel.de/notes/quiet-systems/A cron job that has run without incident for three years earns more trust than any dashboard. I tend to prefer infrastructure I don't have to think about every day. I've traced plenty of incidents back to additions that were meant to help: extra logging or alerts that brought their own failure modes. More instrumentation doesn't necessarily tell me more about what's wrong.Tue, 02 Jun 2026 00:00:00 GMTsystemsfieldhttps://facko.decortel.de/notes/automation-as-leverage/https://facko.decortel.de/notes/automation-as-leverage/A script that saves twenty minutes a day is more useful to me than being known as “the automation person.” The script doesn't need defending in meetings. I start with repetitive work I understand well and can undo if something goes wrong. Judgment calls stay with a person until there is a good reason to trust the system with them.Mon, 18 May 2026 00:00:00 GMTautomationleveragehttps://facko.decortel.de/notes/private-infrastructure/https://facko.decortel.de/notes/private-infrastructure/Moved another service off a third-party SaaS this month and onto the homelab. I prefer being able to inspect the setup when it breaks. That doesn't mean I always know why it broke. I give up some convenience and take on the maintenance. In return, I can follow a problem through the system myself instead of waiting in a support queue.Thu, 30 Apr 2026 00:00:00 GMTself-hostingprivacyhttps://facko.decortel.de/notes/ai-workflows-without-hype/https://facko.decortel.de/notes/ai-workflows-without-hype/Treating a language model as a junior collaborator with no memory and occasional confidence problems has been more useful than treating it as an oracle. It drafts, it doesn't decide. The workflows that stuck: first-pass summarization, boilerplate generation, and rubber-duck debugging. The ones that didn't: anything requiring the model to know something true about the world today.Fri, 10 Apr 2026 00:00:00 GMTaiworkflowshttps://facko.decortel.de/notes/digital-hygiene/https://facko.decortel.de/notes/digital-hygiene/Rotated credentials, pruned OAuth grants, and read through a year of third-party app permissions I'd forgotten I'd approved. Half of them belonged to services that no longer exist. Nothing much to show for this. The permissions list is shorter, which was the point.Sun, 22 Mar 2026 00:00:00 GMTprivacyhygiene