I develop AI systems that belong to whoever runs them.
AI should not mean handing knowledge and control over to a vendor. That is why I focus on how to develop AI systems that work with your own knowledge while remaining as independent, controllable and traceable as possible.
I lead the AI focus group at NOVEDAS Consulting (COHEMI Group). Day to day, I work as a Solution Architect for cloud and interface architecture.
Whatever else I develop in the field of AI, I publish on GitHub, along with the architecture decisions behind my projects.
It rarely fails because of the model itself. The real challenge is accessing the right knowledge at the right time. I develop systems in which agents can find, contextualise and use this knowledge in a targeted way. This is about more than pure similarity search. It is about keeping your own knowledge where it belongs.
For me, AI is not only about what is technically possible, but also about where data is processed and who stays in control of it. That is why, where it makes sense, I rely on local models and European infrastructure. Data protection, the GDPR and the requirements of the EU AI Act are considered from the architecture stage onwards, not as an afterthought.
A working prototype only becomes a good system once it also runs reliably and traceably in operation. I develop architectures for agentic systems that can be orchestrated, monitored and evolved. Even later on, how and why a system reached a decision remains transparent.
Since 2004, I have been working on software, system and cloud architectures across various industries and technical environments. My focus is on scalable cloud architectures, interfaces and data models. Complex systems are not only built but also clearly documented, for example with arc42.
An AI agent reads a Markdown manifest with eight questions and eleven rules and builds a complete enterprise knowledge store in a single session. Three layers, clear write permissions, and tested governance. No vector database as the foundation and no cloud dependency. This article explains the architecture and what it means for operating AI systems in Germany.
Read the article →Alongside consulting, I build my own software. Created privately, documented publicly. Anyone interested can read in the repos how the architecture decisions came about.
File-based knowledge system for AI agents. Three layers, five operations (Ingest, Query, Lint, Review, Eval). No vector index is needed.
MCP server for secure, structured agent access to Obsidian vaults, including graph analytics and a vault health check.
Detects AI-typical patterns in German text: 40 patterns in 7 categories plus a stylometry scanner for sentence-length variance, filler phrases and typography.
AI systems and the architecture underneath are my subject. I am happy to talk about them any time, easiest by email or on LinkedIn.