AI Strategy & Enterprise Architecture
Use-case portfolios, target architectures, technology decisions and roadmaps from proof of concept to platform.
- Architecture Assessments
- Build vs. Buy
- Blueprints & Standards
Enterprise Architecture · Industrial AI · Sovereign Systems
AI systems with substance. From architecture to operation.
I design and build sovereign AI, data and agent architectures for industrial and mission-critical applications — traceable, integrable and close to production.
Before you read on
It runs here. Publicly, at any time.
ZenFactory is not a presentation but a running system: five plants, real OPC UA, a machine from 1998 over Modbus — with a chain of evidence and disclosed limits. Everything stated on this page, you can check there.
5 plants in continuous operation21.8 million rows of history29 assets with Asset Administration Shell207 automated testsEvery intervention in the evidence ledger
Technology offering
The architecture that turns industrial data into productive AI systems.
Not isolated tool consulting: I connect target picture, platform, data, models, integration, security and operations into a robust overall system.
Sovereign AI Infrastructure
Local and hybrid inference platforms on your own GPU infrastructure. Sovereign here means verifiable: your own data, inference on your own network, controlled model delivery and traceable execution — instead of a dependency that can’t be unwound later.
- Multi-GPU Inference
- Model Routing
- Containerised Runtime
Agentic Systems
Orchestrated agents and tool integrations with explicit roles, permissions, states and termination criteria.
- Agent Orchestration
- MCP & Tooling
- Human-in-the-loop
Industrial Data & IT/OT
Data flows and integration architectures between machines, edge, platforms and enterprise systems. See a live example →
- Industrial Data Pipelines
- OPC UA & Interfaces
- Asset Administration Shell (AAS)
AI Governance & Guardrails
Controls, evidence and operating limits that embed probabilistic AI in accountable processes.
- Policy & Safety Gates
- Audit Trails
- Drift & Quality Signals
Platform Engineering & Operations
Production-grade environments with automated deployment, observability and clear operating models.
- Linux & Containers
- CI/CD & Automation
- Metrics, Logs & Traces
Technology stack
From the machine to the model.
AI Workflows · Decision Support · Web Applications · APIs · Human-in-the-loop
Local LLMs · Multi-Agent Orchestration · RAG · MCP · Tool Calling · Evaluation
Python · SQL · PostgreSQL · ETL/ELT · Vector Search · Event & API Integration
IT/OT · OPC UA · Industrial Data · Edge-to-Cloud · Digital-Twin Architecture
Linux · Docker · GPU/HPC · Local-first Infrastructure · Automation
Governance · Safety Gates · Auditability · Observability · Drift Detection · DevSecOps
Technologies are building blocks. What matters is how cleanly data, models, tools and responsibilities work together at the system boundaries.
Reference · publicly accessible
A system that runs. Walkable, not described.
ZenFactory is a virtual factory serving as a reference for industrial IT/OT: five plants, real OPC UA — and a machine from 1998 that only speaks Modbus. Both end up in the same digital twin.
Note: the ZenFactory reference pages are currently available in German only.
OPC UA & Modbus/TCPSparkplug BTimescaleDB historianMAPE-K with approvalSigned evidence
The control loop may change setpoints — but only after four instances in turn have had the chance to say no, and one of them was a human. Every intervention is recorded in the evidence ledger with its rationale, approval and measured effect.
Getting started
One plant, two weeks, one number you can rely on.
A first engagement has to be small, well scoped, and its outcome known in advance. That’s why I always start the same way — regardless of how big the target picture becomes later.
- 01Connect what’s thereOne plant, its controller, read-only. No intervention in the automation.
- 02Measure instead of guessTwo weeks of history, and from it the loss calculation — with your assumptions, not mine.
- 03Only then decideWhether the expansion pays off is then stated in euros, not in a presentation.
Delivery
From business problem to running system.
- 01OrientUnderstand the problem, process and responsibility
- 02FocusClarify benefits, risks and use cases
- 03ArchitectDesign data, integration and boundaries
- 04BuildBuild a production-grade vertical slice
- 05SecureIntegrate governance, monitoring and operations
- 06ScaleEnable teams with blueprints
Stance
What I believe in.
The bottleneck is not the model.
Models are the easiest ingredient today. What holds projects up is whether anyone can take responsibility for their output. Whoever builds the chain of evidence first can swap the model later at any time — the reverse doesn’t work.
Autonomy is the easy problem.
A system that intervenes on its own is quickly built. The proof that its interventions were better than what would have happened without them is the real work — and without it, no plant manager will sign off on a release.
Every model needs a way back.
A model may be a supplement, never a dependency. If it fails or stays silent, the rule that exists anyway must take over. That costs a day to build and decides whether a plant ever puts it into operation.
These aren’t convictions from conference talks but consequences of systems that were built. Where they can be checked, the path there is laid open in the reference plant — including the places where something didn’t work.
About me
Industry understood. Systems built. Transformation led.
My background combines mechanical engineering and industrial processes with software, data, aerospace, manufacturing, product management and enterprise architecture.
Today my focus is industrial AI systems and sovereign infrastructure. I work at both the management and architecture level and hands-on with the system. That yields solutions that fit the business, are technically feasible and remain controllable in operation.
“I don’t just translate between business and IT. I join both sides in an actionable architecture.”
Industrial FoundationSoftware & DataAerospace & ManufacturingEnterprise ArchitectureIndustrial AI
Frequently asked questions
What people usually ask before a first conversation.
How does a collaboration begin?
With one plant and two weeks. It’s connected read-only, without intervening in the automation; afterwards you have a loss calculation built on your assumptions. Only with that number does it become clear whether a larger project makes sense at all.
Do you consult, or do you build?
Both, and that’s the point. I lead at the architecture and management level and work on the system at the same time. An architecture that has never been tested against running code rarely holds up in operation.
Does AI have to go to the cloud?
No. For most industrial applications a server on your own network is enough — with the advantage that data doesn’t leave the building and costs stay predictable. Where a cloud makes sense, I’ll say so; it’s just less often the case than assumed.
Will a model replace our existing rules?
It sits alongside them. A model delivers a finding in the same form as a rule; if it stays silent or fails, the rule answers. That way machine learning remains a supplement and doesn’t become a dependency.
How do I know your claims hold up?
By the running system. The reference plant is publicly walkable, every metric is traceable to the individual measurement, and the limits are stated right there — including the accuracy a prediction model still achieves on a noisy signal.
Contact
Which system do you want to make possible?
For architecture, technical strategy, AI platforms and production-grade implementation.