I didn’t set out to build an AI-driven workflow. It started as a simple experiment — running models locally just to see what I could get working without cloud APIs. Nothing serious at first. Just scripting help, file organisation, small automation tasks.
Then it slowly stopped being “just a tool”.
Now it’s something I actually rely on.
What started as a side experiment has turned into a local AI agent system that sits inside my workflow — handling development, infrastructure, and day-to-day automation across my setup. Built around a homelab, an RTX A5000, and GLM-5, it’s become less about “asking AI for help” and more about letting it actively run parts of the system.
The next step is where things get more interesting.
Moving Toward an Always-On System Layer
I’m moving these agents toward an always-on system layer inside my homelab — not just reactive, but continuously aware of what’s happening. Monitoring services, watching for failures, and keeping systems stable without constant manual intervention.
That means:
- Automated health checks
- Dynamic Docker container management
- Routine server maintenance handled in the background
The goal isn’t just automation — it’s removing the need to constantly think about whether everything is running correctly.
If something breaks, I want it detected early. If something can be fixed safely, I want it handled immediately. If it needs attention, I want it surfaced clearly instead of buried in logs.
Development Side
On the development side, it’s starting to feel less like using an assistant and more like working with an additional layer in the stack.
- Refactoring codebases
- Handling dependency updates safely
- Spinning up isolated environments for testing changes before anything goes live
The direction here is simple: reduce the friction between writing code and shipping it.
Eventually, I want this to feel like a collaborative system rather than a tool I call when I get stuck.
Beyond Development
Outside of development, I’m also pushing into general workflow automation:
- Email handling
- Document summarisation
- Connecting tools so I’m not constantly switching contexts just to get simple things done
The long-term goal is pretty clear: strip out repetitive cognitive load wherever possible so I can focus entirely on building, experimenting, and iterating faster.
TLDR
I started running AI models locally (RTX A5000 + GLM-5) as an experiment. Now they’re embedded into my workflow for file management, automation, and development support.
Next step: turning them into an always-on system layer for my homelab and dev environment — something that monitors, maintains, and actively runs parts of my setup with minimal input from me.
← Back to research archive