AI systems
A model can now make the decision a person makes when they read a document. We build the system around it with you, so it runs in production with the cost measured and the errors caught.
Almost every company has a process with a person in the middle. Orders, claims or requests come in as email or PDF, someone reads them, decides what they mean, and enters the result somewhere else. A language model can make that decision now. When the software around it can act on the decision too, that’s what people call an agent.
The catch is that all of it looks good in a demo. A model gives a plausible answer to almost anything on the first try, and the demo and the production system look the same. The difference is everything around the model: how you test it, what happens when it fails halfway, what it costs per document. That’s ordinary engineering, and it’s what we do.
- Claude
- Gemini
- Mistral
- RAG
- TypeScript
- Node.js
- OpenTelemetry
- SigNoz
- GitHub Actions
Something comes in
an order, a claim, a request
Someone decides
the step that needed a person
Something gets done
a record written, a case routed
The two ends have been automated for decades. The middle is the part a language model does now.
What we do.
You know AI could do something for you, and nobody has worked out what.
Before anyone writes code, we go through your processes with you and the people who do the work. We look for the steps where a person reads something and decides. Then we work out where a language model is worth using and where it isn’t. You get a plan: what to build first, and what it will cost.
You don’t know which model to use, or what it’ll cost.
Extracting, classifying and summarising need different accuracy and cost different amounts per token. We route each task to the model that suits it and measure the result. A cheap model handles what it can, and the expensive one is used only where the accuracy is worth paying for.
Your data is in emails, PDFs and scans.
Most of what’s worth extracting was written for people, not machines. The right answer starts with finding the right part of the document. So we split each document along its own structure, and we test against records where you already know the answer.
You need to know what this costs to run.
Token spend moves with traffic like any other cost, so we instrument it like one. You see cost per document and per model, next to latency and error rate.
You want an agent.
An agent is a loop around a set of tools, and the loop is the easy half. What matters is which tools it can call, what each one is allowed to do, and what happens when a step fails halfway. We build them so a failed run leaves your system exactly where it started.
Tell us what you’re building.
Our clients are in the DACH region and across the EU, and we work with all of them remotely.