Consulting · Automation · AI in your own house
Automation that runs on your own hardware
I automate recurring processes in SMEs — with n8n and language models that run in your house instead of at OpenAI. No subscription you cannot cancel. No data that leaves the building.
I come from infrastructure, not from consulting. Which means: I do not build anything that only works in the demo, and I do not hand over anything that stops working without me.
How to tell you are in the right place
If at least one of these sentences sounds familiar, a conversation is worth your time.
- "Someone here does that by hand every Monday." Pulling numbers out of one system, typing them into another, sending an email afterwards. Four hours a week that no one books as a project, because they are spread across twelve people.
- "When my colleague is on holiday, it just stops." A process hangs on one person and their head. It is written down nowhere.
- "Our data protection stopped the tool." The provider would have been done, but the data would have left the house. Nothing has happened since.
- "We pay three subscriptions for things we could host ourselves." Per user, per month, a bit more expensive with every price increase.
- "We tried AI, but nothing made it into production." There was a pilot. Afterwards nobody ran it.
Why I put this on your hardware
Most of the automations offered to you today are interfaces to American APIs. That works — until one of three questions comes up.
Where does the data live?
As soon as a workflow makes a decision ("is this email a complaint?"), content goes to a model. With a cloud API that means: customer data, quotes, personnel files leave the house. If the model runs at your place, the question is not answered but moot.
What does it cost in two years?
Cloud automation costs per run and per user. The price rises exactly when the solution starts to pay off. Your own hardware costs once and then electricity.
What happens when the provider stops?
n8n on your server is yours — including the workflows. If you no longer want me, it keeps running. That is built that way on purpose.
Where the cloud is the better choice, I say so too. There are cases where your own hardware does not pay off. You recognise them most honestly in the check.
The three packages
Three levels. You start small, at a fixed price, so that a yes costs you nothing you might regret.
Automation Check
The entry point. Fixed price, fixed scope, no obligation to continue.
€1,200 net
Fixed price, no post-calculation.
Who it is for
You know that processes eat time, but not which one pays off first — or whether any does.
What happens
- A video call of about two hours. You show me three to five processes, I ask questions until I understand them.
- I look at which systems are involved and how they can be connected.
- You get a written document.
What you have in your hands afterwards
- Every process assessed individually: effort for the automation, estimated savings, risk.
- A clear order — where to start, what can wait, what I advise against.
- A sketch of what the target architecture looks like at your place: what goes on your hardware, what does not.
- An assessment of data handling that you can put in front of your data protection officer.
- A reliable order of magnitude for the implementation.
Duration: appointment within two weeks of commissioning, document five working days after the call.
Explicitly not included
No implementation, no access to your systems, no installation, no procurement of hardware or licences, no review of your IT security, no legal opinion.
Credit: If you commission the implementation within three months, the amount is credited in full. The check then costs you nothing.
First workflow in production
The usual first project.
from €6,000 net — fixed price, scope defined in the check.
Only once it is clear which process, which systems and which interfaces, does the number stand. But then it is fixed: what was agreed in the check costs what the offer states. Before that I name no price I cannot keep.
Who it is for
A process is identified and should run — not as a pilot, but in day-to-day operation.
What I do
- Set up n8n on your server, or use your existing instance.
- Map one process in full, including the connection to up to three systems you already have.
- If the process contains a decision: set up a language model locally for it and test it on your real cases.
- Monitoring and notification, so you notice when something hangs — before a customer does.
- Documentation and a briefing for your team.
What you have in your hands afterwards
A process that runs without me, on your hardware, with documentation your admin understands.
Duration: four to six weeks of calendar time.
Explicitly not included
No more than one process, no procurement of hardware, no development inside your line-of-business application, no data migration, no taking over your running IT operations, no on-call.
Operations and expansion
For afterwards.
€1,200 net per month for 8 hours.
Who it is for
The first workflow is running and you want neither to run it yourself nor to start from zero again on the next one.
What is included
- A fixed hour allowance per month: 8 hours.
- Maintenance and updates of the existing workflows and the n8n instance.
- Adjustments when a connected system changes.
- New workflows from the allowance, when maintenance does not use it all up.
- A fixed contact person instead of a ticket system.
Duration: three months minimum term, afterwards cancellable monthly.
Explicitly not included
No guaranteed response time outside the agreed times, no on-call on weekends, no responsibility for your core systems, no standing in for your IT.
Need something in between or besides? Individual days are possible, for example for a second opinion on a planned architecture. Day rate: €1,100 net.
What I do not do
So you do not have to ask:
- No chatbots without a problem behind them. If AI is on the agenda and there is no concrete process behind it yet, I tell you in the first conversation instead of selling you a project.
- No model training. I operate existing models on your hardware. Training your own models is almost always the most expensive and worst solution.
- No employee monitoring. Automations whose purpose is performance monitoring I do not build. That is not legal advice, but a decision.
- No legal or data protection advice. I tell you where which data lives and how it flows. The legal assessment is your data protection officer's job.
- No on-call. See capacity.
- No taking over your IT operations. I build and hand over. Operations only within package 3 and only for what I built.
Work samples
I have no client logos to show — I work part-time and so far have mainly built on my own infrastructure. What I can show is how I work. These three systems run in production at my place.
Server operations in chat instead of over SSH
Problem. Whoever runs a server alone loses most time not on the problem, but on checking: logging in, checking logs, applying updates, restarting a service. In the evening, on weekends, in between.
Approach. n8n monitors the machine and applies updates. A Telegram bot is the interface: messages come into the chat, commands go back out of the chat. Managing the Minecraft server also runs over it — status, start, stop, without an SSH session.
Result. Looking at logs and restarting a service is a short chat command instead of a login round over SSH. What used to be a session at the computer, the bot does on request.
Transferable to. Approvals, incident reports, status queries — everywhere someone opens a system today just to look.
Four services on your own hardware instead of four subscriptions
Problem. Nextcloud, GitLab, n8n and a vector database normally mean four providers, four invoices and data in four places.
Approach. One machine, bare metal behind nginx, set up and operated myself. No panel, no managed hosting. Certificates, backups and updates run over the same automation as everything else.
Result. All four services run on my own hardware. No four providers, no four invoices — and the data does not leave the house.
Why this is here. This is the same setup I put in place at your place — with the difference that you do not have to set it up yourself.
Language models that do not leave the house
Problem. As soon as a process contains a decision, it needs a language model. Over a cloud API that means: content goes out, and the costs grow with every run.
Approach. The models run locally on my own hardware, not over third-party APIs. Which one works is decided by day-to-day use, not the datasheet — tested in my own n8n workflows, where the models make the decisions.
Result. Decisions in the workflow are made by the models locally — without a cloud API, without content leaving the house.
How working together goes
- Conversation — 30 minutes, free. You describe what eats time. I ask questions until I understand it. At the end I tell you whether I am the right person. If not, it cost you half an hour.
- Check — two weeks to the appointment, five days to the document. You get in writing what can be automated, roughly what it costs, and what I advise against. Fixed price, no obligation afterwards.
- Implementation — in steps, each usable on its own. No project that only delivers after six months. After every step there is something that runs.
- Handover — documentation and briefing. So you can operate it yourself. Whether I stay on board afterwards is your decision, not the contract's.
What I can manage time-wise
I work full-time as a programmer and infrastructure administrator. I take on projects on the side. That has consequences you should know beforehand:
- I take on one to two projects at a time, no more.
- I answer your enquiry within two working days.
- There is no on-call. If you need someone who is on site at eight in the morning, that is not me.
In return you get someone who has been operating the systems he sets up for you for years — and who does not send you a junior, because there is none.
Questions from inside the company
With a project like this, rarely one person decides alone. So here is what the others will ask.
Your IT lead asks: "What hangs on us afterwards?"
n8n, the workflows and the models run on your hardware and belong to you. You get the workflow exports, the configuration and documentation written on the assumption that I am not reachable. No proprietary format, no licence tied to me.
Your management asks: "What does it cost and when does it pay off?"
The first step costs €1,200 net and is credited in full if you then commission the implementation. The implementation starts at €6,000 net as a fixed price, whose scope is defined beforehand in the check. Whether it pays off is in the check as a calculation, not a promise: hours times frequency against one-off effort. If the calculation does not work out, it says so just the same.
Your works council and data protection officer ask: "What happens to the data and to the people?"
The data stays on your systems — that is the core of the approach, not a concession. For every workflow you get a list of which data it touches and where it flows. Automations for performance or behaviour monitoring I do not build. If co-determination is required, you clarify that internally — I deliver the technical description you need for it.
How we start
Do not write me "I would be interested". Instead answer the six questions below — it takes five minutes and saves us both a conversation that does not fit.
Prefer to write first? or LinkedIn. I will still happily answer the questions above together with you.
By the way
On the blog I write down what I learn while building — n8n, local models, server operations. No signup, no strings attached. → Blog
Besides that I run itorwhat.com, a shop for T-shirts and other print-on-demand items. Nothing to do with the consulting, but the question comes up.