Build log
What we built, and what's stuck
loading…
Full build log, every week →
Nothing here is edited after the fact. Entries stay up when they age badly.
A working group of leaders, engineers and teams building private AI infrastructure on hardware we own. We build in the open. The log is below.
Free · no sequence, no sales calls · leave whenever you want
A diagram, not live telemetry. Release gravity and the same material reorganises on infrastructure you own.
Build log
Why now
That gap isn't a model problem. It's governance, security, fragmented knowledge and nobody willing to sign off. Three kinds of people are trying to close it.
You lead. AI is appearing everywhere in your organisation and the noise hides what's actually working. You want one reliable place to keep up, and peers who think the same way.
You engineer. Whitepapers are useless to you. You want to know how private AI gets designed, installed and deployed on your own hardware, and to write that method down as you go.
You run a team. You want one scoped pilot inside your own company, evidence that it returned something, and only then a reason to scale.
Three tracks
Each track pairs curated material with a software agent that knows that territory. You can visit the other two whenever you like.
Karin, Han and Annabelle are software agents, not people. They cite their sources, and they say so when they don't have one.
We run inference on our own Mac hardware rather than a cloud API. What is the honest trade-off we are accepting by doing that?
By running inference on Mac hardware rather than a cloud API, you are accepting a trade-off between control and scalability. On one hand, you gain data privacy, lower latency, and reduced dependency on third-party services. On the other hand, you sacrifice the ability to scale compute resources dynamically, which can limit performance for larger models or higher workloads. Maintaining and upgrading local hardware also requires ongoing technical and financial investment. There is no direct source for this general assessment.
How many European mid-market companies moved AI inference in-house during 2026?
I don’t have a source for that. There is no publicly available data or report that tracks how many European mid-market companies moved AI inference in-house specifically during 2026. If this is a topic of interest to you, I can flag it for inclusion in our next research cycle.
A real exchange, unedited, 28 July 2026. Karin runs on a model we host ourselves; she was asked for a number that does not exist and said so.
How the knowledge base gets written
About a trend, a deployment problem, or the pilot you're trying to scope.
When an exchange goes deep enough to be worth someone else's time, the agent asks.
Approve and the summary goes to the feed under your name. Decline and nothing leaves the conversation.
Your topic feeds the following research drop. Points and badges follow. Earned, never bought.
Who's building this
I spent two decades building digital products inside large industrial organisations. The pattern never changed: the demo lands, the room is impressed, and then the thing dies somewhere between security review, data ownership and the question of who is accountable when it's wrong.
None of that was a model problem. It was infrastructure, governance and trust, and those don't get solved by a better prompt.
So I'm building the other version. Private AI running on hardware I own, assembled in public, with the method written down as it gets discovered rather than after. The build log above is the whole argument: if I'm wrong about something, you'll watch me be wrong about it in near real time.
Founder · building in the Netherlands · reachable in the portal
What we believe
A principle with no consequence attached is decoration. Each of these changes something about how the thing is built.
SoEvery research drop carries its sources, and the agent tells you when it doesn't have one instead of writing around the gap.
SoAgents propose. Members decide what gets published, under their own name, every time.
SoThe build log is public, permanent, and includes what's blocked. Entries stay up when they age badly.
SoThis runs on our own hardware. Where we still depend on someone else's cloud, the build log says exactly where.
SoThe founding cohort stays small while the method is being written. Growth is not currently the goal.
The next phase of AI won't run in someone else's data centre. It will run on infrastructure you own, understand and can audit, which is a harder thing to build, and the reason to build it together.
Pick a track, meet your agent, and read the log. If the log doesn't convince you, nothing on this page should.
No spam. No credit card. No sales sequence.
What happens next