AI engineering: MCP, models and documents.
The technical layer the other services stand on: MCP servers, model selection and fine-tuning, document processing at scale, and software built agentically — under the control of architecture, specifications and tests.
This is where it is usually decided whether an AI deployment will be maintainable. Which model, how it is given context, what a single call costs, and what happens when the provider changes the API — those decisions are made once and paid for over the following years.
I built n8n-mcp, one of the most widely used MCP servers, and fine-tuned my own model for generating n8n workflows. The same tools and the same approach go into client projects.
What I do
MCP servers
I design and build MCP servers that give AI agents controlled access to your systems — with validation, permissions, and verification of the agent's work before it goes any further.
Model selection and fine-tuning
I match the model to the task and, where it pays off, fine-tune a custom one. The aim is usually the same: frontier-level quality at a fraction of the cost, without sending data outside.
Document intelligence and RAG
Extraction and structuring of data from documents — PDFs, scans, mixed formats — and semantic search over very large collections. Scale measured in hundreds of thousands of documents.
Agentic software development
Applications and backends built agentically with Claude Code and MCP, always under the control of architecture, specifications and tests. The agent writes the code; whether the system is built correctly is decided by oversight.
Architecture, cost and security
A review of an existing solution for call costs, data leakage, single-vendor lock-in, and what happens at ten times the traffic.
Who this is for
- Product teams shipping AI features who need an architect with production experience.
- Companies whose API bill is growing faster than the benefit of the deployment.
- Organisations sitting on a large document collection they currently cannot query.
- Teams that want to adopt agentic software development without losing control of quality.
How I run it
- 01Review and architectural decisions
I start with the decisions that are expensive to reverse: the model, how context is supplied, system boundaries, where data lives.
- 02A prototype measured in numbers
I build a version that can be judged — accuracy, cost per call, latency. Only then is there anything real to say about scaling.
- 03Production
Deployment with monitoring, error handling and cost control. Without those, the first bill or the first outage ends the project.
- 04Maintenance
Models and APIs change every few months. I stay so those changes aren't a surprise.
What this is based on
Own projects where this layer is the product — and client work where it sits underneath.
The tool that lets AI agents build working n8n automations — 22,600+ GitHub stars, 128,000+ users and 1.1M+ workflows built, including at Deutsche Telekom, PayPal, Mercado Libre, MIT and NYU. Built entirely agentically.
Read the case studyTwo World Bank workstreams in Romania's energy sector — ~100,000 heterogeneous energy certificates (photos, OCR PDFs, spreadsheets) consolidated into one Postgres database with heavy LLM extraction, and the 2021 census turned into accurate heating maps across all 42 counties. Delivered ahead of deadline.
Read the case studyCustom-trained models that build n8n workflows from a plain-language request — frontier-level quality, self-hosted, about a cent per workflow.
Read the case studyA research project in a field I knew nothing about — 3D modelling, from zero to an interactive browser app in three days. A test of whether the method holds up outside my own expertise.
Read the case studyCommon questions
- What is MCP, and is it just another trend?
- MCP is the standard that lets an AI agent use external systems in a controlled way, instead of hard-wiring a separate integration into every tool. I built one of the most widely used MCP servers, so I speak about it from the maintenance side rather than the announcement side.
- When does fine-tuning make sense, and when is a good prompt enough?
- Fine-tuning pays off on a narrow, repetitive, high-volume task — it cuts cost and lets you run the model yourself. For variable tasks or small scale, a good prompt and the right model choice win; tuning would only make changes harder.
- Can this run inside our own infrastructure?
- Yes. Some clients require that data never leaves their environment, in which case I deploy locally or into your cloud. It affects model choice and cost, so we settle it at the start rather than the end.
- Code written by agents — how is quality controlled?
- The same way as with a team of people, only faster: specification before implementation, tests, code review and multi-agent CI. Responsibility for the architecture, and for what reaches production, stays with me.