The Homemade Machine is my personal research project on home-built AI systems, practical home automation, local LLMs, fine-tuning, and agentic workflows, published through the same agentic content pipeline the project is trying to improve. This is a 12-month research project. Full disclosure: I am not a trained researcher or data scientist, and that is part of the learning process. The posts are about work I am personally exploring, not abstract AI commentary. This site is both the publishing surface and the experimental substrate. The subject matter sits around home-built AI systems, practical home automation, local LLMs, fine-tuning, agents, tools, models, and the craft of making machines useful.
Who Runs It
This site is run by me, Craig Foster, and all posts relate to my personal explorations on AI. This is not in any way related to my full-time employment. It follows the systems I am building for personal use, or experimenting with, and where I am learning how to make those systems more reliable. I am based in the UK and work in the technology sector as an HR IT product manager, so the use of AI and the impact of AI on knowledge work is something I have both a personal and professional interest in. But I am not an engineer; my background is in operations, so part of this work is also about personal capability growth with these weird, new, agentic systems.
What Gets Published Here
The posts are not written from a distance. They come from my personal workbench: local LLM setup, model serving, practical evaluation, home automation experiments, fine-tuning, adapters, model-routing decisions, and agentic systems I am trying to make dependable. I expect to publish reviews of hardware, systems, tools, and models, alongside analysis, benchmarks, and implementation notes from the actual work. The useful line for me is whether something holds up under real use, not whether it sounds impressive in the abstract.
What The Experiment Is Measuring
The core question is not "can AI write a blog post?" It is whether an agentic publishing pipeline can keep producing useful work over time while its quality, relevance, cost, and failure modes are measured honestly. I am tracking how much human intervention is needed before a draft is publishable, how often published posts need corrections later, how much each clean post costs to produce, whether verifier systems agree with human judgement, and how long approval actually takes. The aim is to reduce unnecessary human-in-the-loop work without pretending that lower effort counts as progress if the output gets worse. All of the published articles will be based on actual work with real examples, not just AI theoretically writing about AI.
The Publishing Pipeline
The publishing pipeline is part of the experiment, not just a tool behind the site. It helps draft, route, review, and publish work under human supervision, with the goal of continuously improving the quality of the content while reducing avoidable corrections, cost, and human-in-the-loop effort where the data supports it. Human approval remains part of the system until the measurements justify changing that. The learning record is also part of the output: I want to publish what I learn while the machine that helps publish it is being improved.
The Two Streams
Posts are labelled by stream so readers can see what kind of work they are reading. Automated-stream posts are AI-drafted, human-reviewed before publication, and included in the experiment's autonomy and cost metrics. Learning-stream posts are written by me and are used to document what is happening in the experiment, what failed, what changed, and what I am learning along the way. They are deliberately excluded from the automation metrics, because the point is to keep the measured machine output separate from the human record of the work.
Editorial Standards
AI may help draft some posts, but I remain responsible for what gets published here. Benchmarks, screenshots, and recommendations should come from first-hand measurement, not invented filler or second-hand confidence. If a link is commercial or may generate commission, it should be disclosed. If something is wrong, it should be corrected. And when I write directly about the learning process, those posts should stay separate from the measured automated-stream performance. The Disclosure and Privacy pages cover the reader-facing details.
Closing
The useful result after 12 months is not a perfect machine. It is a clear record of what worked, what failed, what stayed relevant, what got cheaper, what still needed human judgement, and what that says about building autonomous systems as a knowledge worker without being an engineer.