Independent IT consultant in Hamburg, Germany — interim IT management, project leadership and strategy work for mid-sized companies.

For about a year I have been running my own business on agents: bookkeeping, document filing, inbox triage, editorial work, changes to my own website. This account is where that tooling ends up, one repository at a time, together with the reasoning that made me build it that way. The code is cheap and reproducible. The decisions are the part worth publishing.

These tools touch real business data — mail, invoices, client context. So the constraints came first, and they live in the code rather than on a policy page:

• Measure locally instead of uploading. Where a model only needs numbers, it gets numbers. housestyle learns a writing voice from a corpus that never leaves the machine.
• Data that isn't needed stays out. Corpora, profiles and anything carrying a client's name sit outside the repository — not one .gitignore line away from being committed.
• Access to production systems is tiered. Read, diagnose, draft, live: each step its own permission, with a mandatory backup and a log before anything is written.
• Agents prepare, people publish. Automating the preparation is the win; automating the decision to publish is not. Nothing goes out without a person approving that specific piece.
• Every release gets a security pass. No absolute paths, no mailboxes, no domains, no client references in what ships.
• Apache-2.0, for the patent grant.

None of this is clever. It is written down and actually enforced, which in my experience is the part that usually isn't.

• efficas.ai — field notes from AI projects: what I built, what it cost, and what didn't work. Currently in build.
• efficas.de — the consulting side.
• LinkedIn

If you set one of these up — whether it worked or not — I would like to hear about it. That is worth more to me than a star.