A small software studio in Europe that builds AI features into real products, and keeps them working.
We build LLM integrations, agents that use tools, and search over a company’s own documents, from the first idea to production.
Small on purpose. We keep the team small so that the people who design a system also build it and answer for it. You talk to the people doing the work, not to an account manager who passes messages on.
A lab, not a factory. We try ideas on our own products first. This website runs on a content system we work on in our lab, and we use AI agents in our own daily work. What we recommend to clients, we have used ourselves.
Measured, not promised. Every feature we build comes with a test set and tracing. When we say something works, we can show the numbers, and when it stops working, the numbers show that too.
For the long run. Many AI features look good in a demo and fail in their third month. We build for the third month: limits on cost, checks on quality, and a handover your team can run.
The green duck is the mallard drake. On the water it looks calm. Under the water, its feet never stop.
We chose it because that is how a good AI feature should work. Your users see a calm, useful answer. Underneath, the feature is tested, measured and watched every day. The “g” in our logo is the duck’s head.
We work from Târgu Mureș, a university town on the Mureș river in Transylvania, Romania.
That puts us inside the European Union, under the GDPR, on Eastern European Time. Our working day overlaps fully with clients in the rest of Europe and the UK, and with the morning in North America. We meet on video, write everything down and work in your repository, so distance rarely matters. When a project needs us in the room, we travel.
We take fewer projects so that each one gets our full attention.
We say yes to a real problem, a good fit for AI, and someone who can decide.
A real problem means real users with a need you can describe in a paragraph. A good fit for AI means the task involves language, documents or many small judgement calls, and a wrong answer can be caught and corrected. Someone who can decide means a person who can answer our questions within a few days and say yes or no to the trade-offs every project meets.
We also look at the data. A feature is only as good as what it reads, so before we agree to a project we want to see a sample of the documents, tickets or records it will use.
The habits behind every project.
Plain answers, early. If something is not going to work, you hear it from us before you hear it from your users.
Language models make mistakes. We design for that: answers show their sources, risky actions need a person’s approval, and the feature says “I don’t know” when it should.
Your code, prompts and test sets live in your repository. Your team learns how to run them. You can stop working with us at any time without losing anything.
We use the least data the feature needs, keep it where you decide, and write down every place it goes.
Every claim about quality, speed or cost comes with the test results or traces behind it.
Some work we turn down, however good the budget. Saying it here saves us both a call.
Green Duck Labs is led by Szabolcs Szántó, founder.
He answers the first email, joins the first call and stays on the project until it is handed over. Write to him directly about what you want to build.
Tell us about the product and the problem. A few lines are enough, and we reply within one working day.
You can also read how we work or see what we build in our lab.