01/12
Architecture
Rapid prototyping
Something to click on by Friday.
A working prototype ends discussions that slides cannot. I build quickly to make decisions possible – and say clearly which parts are prototype and which are ready for production.
In practiceAt Bosch, a self-developed harness application for prototyping agent-to-agent communication without writing a single line of code – with export to MCP or n8n workflow artifacts. Before that: proofs of concept at R+V, Clientfocus and Hewlett-Packard.
02/12
Architecture
Solution & enterprise architecture
Systems that still make sense in five years.
Integration, service APIs, platforms that many teams depend on. Good architecture is mostly about saying no early and writing down why.
In practiceArchitect of one of the two core components of NEO at Bosch; ITSM transformations for KfW, Audi, Volkswagen, Clariant, T-Systems and DB Systel – business solutions on premises and air-gapped, plus interface integrations.
03/12
With people
Pre-sales & discovery
Understanding the problem before selling a solution.
Workshops, stakeholder conversations, technical demos. I translate between the people who need something and the people who build it – honestly, including when the answer is “you don’t need AI for this”.
In practiceAt Bosch: internal trainings. Before that: 11+ years of vendor-side pre-sales across EMEA at Clientfocus, WSP and Hewlett-Packard.
04/12
With people
AI trainings & enablement
Teams that understand AI make better decisions about it.
I explain AI without hype and without fear – for engineers, managers and everyone in between, in German and English.
In practiceAt Bosch Digital: several AI trainings and presentations; co-initiated an AI use-case barcamp. Earlier at Hewlett-Packard: training programs for service management teams across Europe, and live product presentations at the HP Software summits in Barcelona and Hamburg.
05/12
AI, ML & data
Agentic workflows
Language models that work in steps – plan, act, check – instead of just answering.
An agentic workflow lets a model choose its next step: call a tool, read the result, decide again. That freedom is powerful and expensive, so I design where autonomy pays off and where a fixed process is the safer, cheaper answer.
My rule: as much workflow as possible, as much agent as necessary.
In practiceBosch agent harness: workflow design based on JSON, built with FastAPI, Python and Node.js. Bosch NEO: an AI resolution engine that analyzes failed transactions and proposes the next action.
06/12
AI, ML & data
Agent-to-agent communication
Specialized agents that hand work to each other – through contracts, not chatter.
When several agents cooperate, the interfaces between them matter more than any single prompt: who may ask whom, what is passed along, how a failure becomes visible. I treat agent-to-agent (A2A) messages like service APIs – typed, versioned, logged.
In practiceBuilds on years as Service API architect at Bosch: infrastructure, RHEL operations, incident, service request and change management, development and deployment.
07/12
AI, ML & data
MCP architecture & function calling
Giving models access to real systems without handing over the keys.
The Model Context Protocol (MCP) standardizes how models reach tools and data. I design MCP servers and function sets that are small, well described and whitelisted – so a model can only do what the business has approved, and every call can be traced.
In practiceFunction whitelisting in the Bosch resolution engine.
08/12
AI, ML & data
Agent harness development & guidelines
The frame around the model: loop, context, tools, permissions, logging, stop rules.
The model thinks; the harness decides what actually happens. I build harnesses and write the guidelines teams need to build their own consistently – so ten agents in a company behave like one architecture, not ten experiments.
In practiceAt Bosch: a self-developed harness application – agent-to-agent prototypes without code, exported as MCP or n8n workflow artifacts.
09/12
AI, ML & data
Guardrails & safe functions
Deciding in advance what an AI may never do.
A model can be convincing and wrong at the same time. Guardrails make that harmless: whitelisted functions, validated inputs and outputs, human approval for anything irreversible, and an audit trail that explains afterwards why something happened.
A blind spot I take seriously: leaking through reasoning. A guardrail blocks the answer – the user never sees it. But the model’s reasoning may still contain exactly that answer. The reasoning ends up in a log file, another agent reads the log, processes the content and passes it on. Current checks mostly inspect the final answer; reasoning traces and logs are covered only partly or not at all.
That is why I treat reasoning and logs as outputs, too: the same checks, redaction before anything is written, and no agent reads another agent’s raw reasoning.
In practiceWhitelisted functions and Teams / Outlook alerting at Bosch; warning flags in the Evolution Engine.
10/12
AI, ML & data
Data science & machine learning
Understanding the data first – choosing the model second.
Before any model there is data: cleaning, normalizing, finding the signal. For large log volumes I build the foundation underneath: ClickHouse as the analytical database, log vaults for secure long-term storage, and a Parquet transformer that turns raw log files into columnar, compressed files that are fast to query.
On that basis I use classical machine learning where it beats a language model, and neural networks where it does not.
In practiceNeural personality models trained per individual in the Evolution Engine; normalized log analysis at Bosch.
11/12
AI, ML & data
Custom RAG
Answers grounded in your own documents – with sources.
Retrieval-augmented generation lets a model cite your knowledge instead of guessing. I build the retrieval myself – chunking, embeddings, ranking – so it is always clear why an answer appeared and where it came from.
In practiceOwn implementation, not a black-box product.
12/12
AI, ML & data
Local & air-gapped AI
Local or not local – that is the question.
How do you protect company data – and does it make sense to own locally operated systems? After all, a single RTX 6000 card costs around €17,798 (October 2026).
My answer is a clear yes for mid-sized companies that want or need to keep their intellectual property to themselves. But be aware: the basic prerequisite is an analysis of the data. A good solution combines AI with deterministic tools and a very well-designed MCP landscape that connects the models cleanly to the systems. Part of it is GPU and VRAM management that distributes the load sensibly, plus scheduled nightly jobs whose results can be used the next day without any GPU or VRAM.
Technically and at the architecture level, I look after models across their whole lifecycle – plan, build, run: from LoRA fine-tuning to operation on GPU Docker stacks, fully offline if required.
In practiceIn the enterprise, a wide range of models (GLM, Mistral, Claude by Anthropic, GPT, Gemini) are connected through a security layer and used in workflows built with n8n or OpenText. In my own projects, the Evolution Engine runs completely offline, and my content pipeline translates with a local model.
