The Blueprint
What You Will Find Here
Behind the calm atmosphere is a technical system that has to work in reality. Here, I explain what I build, why I chose it, how it behaves when something changes, and what the measurements can actually tell me.
- Why I chose one approach over another
- How the automation is meant to behave when something goes wrong
- Which signals I measure, and what they cannot prove
- What is working, what is planned, and what is still uncertain
Details coming soon
Build Architecture
How the system is physically and logically built, where each part runs, and where I draw the technical boundaries.
Details coming soon
Automation Logic
How control routines, operator actions, and conservative defaults are meant to keep the system predictable when conditions change.
Details coming soon
Monitoring & Proof
What I measure, what I may show publicly later, and where telemetry ends and interpretation begins.
Details coming soon
Build Reference
Reference material will appear here once real build details can answer a useful question instead of merely filling a page.
AI & Project Workflow
How I Use AI Beyond the Browser
I do not use AI as a simple question-and-answer box. I work with it inside a structured environment of project files, rules, specialist agents, and review loops. I call that environment the harness.
- Project Context
- The SSoT documents hold the project's decisions, constraints, and plans. They give AI a real working context instead of leaving it to fill the gaps.
- Research & Design
- I define the goal. AI helps me research options, challenge assumptions, and develop the technical details within the wider project.
- Building the Project
- AI already helps maintain the website, refine the system architecture, and prepare repeatable workflows. The planned video workflow will extend that support into DaVinci Resolve.
- Independent Checks
- Important recommendations are checked against documentation, external sources, or another model. A confident answer is still an answer, not automatic truth.
- My Responsibility
- I choose the direction, make the purchases, record the footage, judge the result, and decide what becomes part of MPB. The tools can change; that responsibility does not.
This is not one-click automation. It is a carefully built way of working that lets one person take on a project that would otherwise require a small team.
Setback · Analysis · Lesson
The Machine That Was Never Broken
A new fanless Shuttle mini PC was meant to become one of the quiet workhorses of this project: the control node that runs around the clock while nothing else has to. I fitted the memory module and the SSD myself — routine by now. I had built up several machines of this project the same way, without a single problem.
This time was different. The installer froze at random points. Under load, the whole machine locked up completely. An overnight memory test made it official: 113 logged memory errors, and the machine reset itself roughly every hour. Not what you want from a device that is supposed to run for years.
So the AI and I did what we always do: analyze instead of guess. I handled the hardware and watched the machine; the AI worked through the logs, built hypotheses, and designed the next test. The decisive one seemed clean: the same memory module, moved into another machine, ran flawlessly — at full speed, zero errors. So the module was fine, and the board had to be defective. Logical. Well documented. WRONG.
The turn came from a simple question I asked almost in passing: does the manufacturer publish a list of approved memory modules? It does. Ours was not on it. The AI had not gone looking for that list on its own, and it named the reason openly: it had anchored on the defect theory and read “works in that other machine” as “works, period.” Compatibility does not work that way. A module and a board either match or they don’t — and that is a property of the pair, not of the parts.
The proof was a test with exactly one variable changed: an approved module, same machine, same test stick. Five hours, zero errors, zero resets. Then the same stress test that used to freeze the machine solid — this time it just ran. The machine was never broken. The combination was.
What stays with me is not the fix but the pattern. Careful evidence can still lead to a wrong conclusion when a clean test answers the wrong question. Routine does not replace the check — the approval list takes two minutes and would have saved days. And this is what working with AI actually looks like here: not an oracle, but a fast, thorough partner that can anchor on a theory just like a human can. The value sits exactly where one of us breaks the other’s anchoring. Setbacks like this are part of the build. Pretending otherwise would be the only real failure.
