01Strategy
AI readiness and use-case selection
Where AI realistically helps in your stack, where it would just be expensive theatre, what GDPR and the EU AI Act mean for it, which architecture you actually want to operate.
/ AI · tool, not religion
Use AI thoughtfully, not blindly
Between hype, reality, and actual value
/ AI services
Where I actually help
Yes, I am critical of AI hype. At the same time, I integrate exactly the systems that genuinely make sense for many businesses. Pragmatic, GDPR-aware, operable.
Common path: a discovery workshop (90 min), then a 1-2 day AI readiness audit, followed by either a concrete integration or an honest verdict that the timing is not right yet.
02Architecture
Local AI infrastructure
Local LLMs (Ollama, llama.cpp, vLLM), vector databases, embedding pipelines, on-prem inference for data that cannot leave: public sector, tax, healthcare.
03Automation
n8n and MCP workflows
Self-hosted n8n pipelines, Model Context Protocol servers, tool bridges to existing systems (Excel, SQL, ERP). AI as a tool between your systems, not as magic.
04Engineering
RAG and document analysis
Retrieval-augmented generation for internal knowledge bases, contract review, inbox classification. Sources traceable, answers citable, hallucinations controllable.
05Agents
Agent frameworks and function calling
Agent frameworks in the LangChain/LangGraph class and function calling, when a task genuinely needs several steps handled autonomously, controlled rather than as an end in itself.
06Tooling
Claude Code, Cursor, Codex in team workflows
How a small team puts Claude Code or Cursor to productive use, which guardrails matter, what code review looks like when half the code comes from agents.
07Governance
GDPR, EU AI Act, operations
Data protection impact assessment, processing agreements with model vendors, EU AI Act classification of your use case, logging, audit trails, accountability. So a demo becomes operations.
08Data
AI in reporting and data prep
AI-assisted support for SQL, ETL, and Power Query, where pattern recognition or code suggestions save real time, without anyone losing track of why a number ends up looking a certain way in the report.
/ Architecture · sample
Local RAG, GDPR-compliant, no cloud egress.
A user question hits a local embedding model, queries a vector DB, and is sent with the relevant internal-document context to a local LLM. The answer comes back with citations, fully auditable, nothing leaves the building.
/ Architecture · stack
Local AI stack in real operation.
Triggers via webhook, cron, email, or an Excel button. n8n orchestrates the pipeline, MCP servers bridge to your real systems (Excel, SQL Server, ERP, filesystem). Inference local, logging and retry built in, no cloud egress.
/ Stance · long read
My stance on AI, in full.
No hype, no panic. Where AI genuinely helps, where it is just expensive theatre, and what actually matters after the first prototype.
AI is neither the salvation of humanity nor the end of every profession. And no, a Claude Code prompt alone usually does not replace stable software, a thought-through process, or years of experience with real systems.
But anyone who completely ignores what AI can do today is already at a disadvantage.
That is why I take a pragmatic view of AI. Where it is sensible, economical, and technically defensible, it can be extremely helpful, for reporting and data prep, internal knowledge systems, or routine tasks. More on exactly where, further down.
But that is often where the real work only begins.
"It works on my machine" is not yet a productive business process.
An AI agent is quick to build. Then what?
Today Claude Code, ChatGPT, or Cursor can, in remarkably little time, build small tools, analyze data, create workflows, write scripts, build automations, connect APIs, and even generate complete mini systems.
And honestly: that is impressive.
But after the first wow moment, the questions that follow are usually far less viral on LinkedIn:
And suddenly you notice: the prompt was the easy part.
AI is not a substitute for understanding
I am not into panic. But I am also not into "AI will replace all developers tomorrow". Reality sits somewhere in between.
AI is a tool. An extremely powerful tool. But a tool is only as good as the person who uses it well.
An AI agent working unchecked on production data is not innovation. It is often just a new failure mode with better marketing.
- personal data and internal company data
- reporting, permissions, and sensitive documents
- customer information
- public institutions
- medical or financial contexts
Cloud or on-prem?
It depends. There is no single best answer for every case.
Cloud (e.g. ChatGPT, Claude, Microsoft Foundry, Gemini)
- Fast to start, strong models, little own infrastructure, straightforward integration.
- Watch closely: privacy, data transfer, running costs, vendor lock-in, dependencies, and sometimes unclear data flows.
On-prem / local (e.g. Ollama, Open WebUI, local LLMs, Qdrant, RAG)
- More data control, can run locally, often more GDPR-friendly, fewer external API dependencies.
- But significantly more complex: maintenance, hardware, updates, security, monitoring, user management, and performance.
Many small teams underestimate how much infrastructure suddenly sits behind "we host it locally."
Public sector and sensitive environments
That is where AI is often far harder than in YouTube demos. Topics like these suddenly matter:
Law and governance
Data protection impact assessment, GDPR, EU AI Act, internal policies, and documentation duties.
Infrastructure and operations
Hosting location, network separation, VPN, logging, and audit-grade records.
Organization
Role models, permissions, and works councils when it comes to rollout and use.
And that is appropriate. A flashy tool is not automatically sensible or permitted in every environment.
Where AI is genuinely useful, in my view
Not everywhere. But often where repetitive work piles up. Typical areas:
Reporting and data flow
Reporting, data preparation, ETL-style processes, and SQL support.
Knowledge and search
Knowledge systems, internal search, support knowledge bases, and sensible summaries.
Office and workflows
Office automation, structured workflows, and small internal tools.
Assistants
Intelligent assistants embedded in existing processes.
Combined with real processes, AI can save a lot of time. But only if:
- data quality is right
- processes are understood
- security is considered
- people stay in control
- systems stay maintainable
What I deliberately do not promise
No magical fully automated companies.
No AI replaces every employee.
No miracle fixes after two prompts.
No agent systems that supposedly do everything alone.
Because reality usually looks different.
My approach
I treat AI as a tool, not a religion. Where it really helps, can be integrated sensibly, and delivers a real economic upside, I am glad to support ideas, evaluation, automation, on-prem AI setups, reporting, workflows, data analysis, internal tools, and office or process automation.
Related topics
/ Next step
Discuss an AI use case?
Briefly describe the scenario, data context, and constraints (cloud, on-prem, public sector). I will reply with an honest view.