AI automation that survives production
Most automation projects do not fail at the model. They fail where the model meets a real system — the database, the CRM, the machine, the API with no documentation. That is the part I build.
What this actually means
Not a chatbot in the corner of a page. It means the manual step somebody has been doing every Tuesday for two years: pull data out of one system, reshape it, type it into a second, check it, chase the missing piece. That step does not disappear because of another tool subscription. It disappears when there is an integration, a data model, and error handling that still holds when the other side answers however it likes.
- Process automation — the path between two systems that were previously connected by a person and a spreadsheet.
- LLM agents with system access — a model allowed to read your own data and call real tools instead of guessing.
- Integrations and data pipelines — API work, caching, real-time data, and an answer for what happens under load.
Agents that do more than reply
An agent only becomes useful once it can reach the systems that hold the answer. There is a standard for that now — MCP, the Model Context Protocol — and I built a server on it that opens Austrian federal law to LLMs: §-level retrieval, historical versions, amendment timelines, BGBl lookup. It is publicly reachable, and the write-up is here.
The same shape fits internal systems: an agent that can read an ERP, a ticket system or an archive properly, through a few well-bounded tools rather than blanket access.
Where these projects actually break
- Data quality before model choice. If two systems spell the same customer differently, that is not an AI problem.
- The failure case. A pipeline without a check produces mistakes faster than a person does. In my own content pipeline a blind judge decides whether anything ships — that is what it looks like built.
- Cost per run. A workflow that costs more per item than doing it by hand is an expensive demo. That can be worked out beforehand.
- Who maintains it. An automation with no owner is a delayed outage. What I build is documented and can be handed over.
How it runs
It starts with a discovery sprint: one week maximum, fixed price, ending in a written scope, an architecture and an honest estimate — with no commitment past it. If the conclusion is that the effort does not pay for itself, that goes in the document too. The three ways of working together are on the home page.
Who this is for
Companies and independents who want to automate something, remote across Europe, in CET. No team in between: a one-person business — whoever hires me talks to whoever builds it.
What is already running
- RIS MCP Server — 8 tools over the Austrian legal information system, live.
- Event-management PWA — HubSpot as the single source of truth, dual-layer cache and dedupe for traffic spikes.
- Industrial control software — a state machine over eight real flush programs, with live hardware diagnostics.
- ClimateCanary — sensor firmware to dashboard across four layers, privacy rule enforced twice.
If there is a process that costs time every week: describe it briefly and I will tell you whether it can be automated and roughly what it takes.