LLM integrations
Chat, search, summaries and writing help inside the product you already have, wired to your data and your users’ permissions.
- Prompt design and versions
- Model choice and routing
- Cost limits per user
We build AI features into real products, and we keep them working once real users arrive.
A software studiofor LLM integrations, agents and search over your own documents.
Târgu Mureș, RomaniaWorking in English with clients anywhere.
Who we are
Green Duck Labs is a small European software studio. We build AI features into real products, from the first idea to production.
Who we are. A small, founder-led studio. The person who answers your first email is the person who designs the system and reviews the code. We keep the team small on purpose, so nothing gets lost between the first call and the build.
What we build. AI features inside products that already have users, or soon will: assistants and chat, writing help, agents that call your APIs, and answers drawn from your own documents with the sources shown. We also build the ordinary software around them.
Where we work. From Târgu Mureș in Romania, inside the European Union, with clients anywhere. We work in English, on European hours, and we meet on video. When your data has to stay in the EU, we build it that way.
How to start. Send us a message with a few lines about the product and the problem. We reply within one working day, usually with questions or a time for a short call.
Two halves of a feature
Above is what your users see. Below is what we build under it. A demo shows only the top half.
Why a duck
We are named after the mallard. On the water it looks calm. Under the water, its feet never stop.
That is how we think an AI feature should work. Your users see a calm, useful answer. Underneath, the feature is tested, measured and watched every day.
That half is the tests that check each answer, the limits on cost, the logs that show what happened, and the alerts when quality drops. We build both halves from the first week, so the feature still works in its sixth month, not only on the day of the demo.
Look under the water
Here is one answer from a document-search feature, the kind we build. Open the trace to see what ran before the user saw a word.
Every feature we ship keeps a trace like this for every request. It tells us why an answer was wrong, what it cost, and whether the last change made things better or worse.
Asked in a staff handbook assistantCan I carry unused holiday days over into next year?
Yes, up to five days. Days above five expire on 31 March of the next year, unless your manager approves an exception in writing.1 Carried-over days are used before the new year’s days.2
1 Staff handbook, 4.2 Holiday2 HR policy update, January
The trace under this answer
- Question checked for personal data Nothing to removePass
- Searched the handbook, by keyword and by meaning 1,240 passages searched6 found
- Re-ranked the results Kept the three that answer the question3 kept
- Answer written, with a citation for each claim Prompt version 14212 tokens
- Each sentence checked against its source Two sentences, two sourcesPass
- Policy filter No advice outside HR policyPass
Time 1.8 sCost $0.0021Logged for review
An example. The handbook, the question and the numbers are made up for this page.
Four kinds of work
Often in one project. Each one ships with tests and monitoring, so you can see how it performs, and most projects mix two or three of them.
Agents
Software that works through a task step by step. It calls your APIs and tools, and hands the case to a person when a rule says it should.
- Tool calling
- Retries and clear states
- Human review steps
Search over your documents
Answers drawn from your own manuals, tickets and contracts, with the source shown next to every answer so people can check it.
- Ingestion and chunking
- Hybrid and vector search
- Checks against made-up answers
Evals and monitoring
Test sets built from real cases and run on every change, plus tracing in production, so a drop in quality shows up before your users notice it.
- Test sets
- Tracing and cost tracking
- Alerts and rollbacks
The software around it
APIs, admin screens, dashboards and deployment. An AI feature needs a normal product around it, and we build that too.
- Node.js and Laravel
- React and Vue
- CI/CD and hosting
The first month
Every project starts the same way, whatever its size. After four weeks you have a working first version, the tests that measure it and a plan for what comes next.
Then you decide whether to go on, with real numbers in hand instead of a promise.
Week one
A call about the goal and the users1 hour
Data and access review2 days
How we will measure successin writing
Week two
Architecture and model choice2 days
First test set from real cases50–100 cases
Safety rulesin writing
Weeks three and four
First working roundshown to you
Test results and costswith each release
The plan for the next roundsagreed together
A working first version, its tests and a plan4 weeks
Built in our lab
This website runs on aicms, a content system we work on in our lab. You edit the site by asking for the change in plain words. An AI agent drafts it, you see a preview, and nothing goes live until a person accepts it.
We use our own tools every day. It is the fastest way to find out where an AI feature helps and where it gets in the way.
- Pages are plain filesEvery page is an HTML file the agent reads and writes, so nothing is locked in a database.
- Every change is a preview firstThe agent proposes, a person accepts, and each accepted change is saved as a commit.
In the agents section, add a sentence saying that every step is logged.
AgentDone. I added one sentence after the first paragraph of “Agents”. Nothing else on the page changed.
Agents, first paragraph
…and passes the case to a person when a rule says it should.
…and passes the case to a person when a rule says it should. Every step is logged, so you can see what it did and why.
DiscardAccept
Ask for a change…
Two ways to work |
A projectOne feature or product, from the first idea to launch. | A retainerA monthly arrangement for features that are already live. |
|---|---|---|
| Good for | A new feature with a clear goal | AI features your users already rely on |
| How it starts | A short call, then a written proposal with scope, timeline and price | The same call and proposal, or the end of a project with us |
| What you get | The feature, its tests, its monitoring and a handover your team can run | A set amount of our time each month for new features, model updates and fixes |
| How you see progress | Working software and test results after every round | A monthly report on quality, speed and cost |
| Price | Agreed in the proposal before work starts | A monthly fee, agreed in the proposal |
| Who owns the code | You do | You do |
| Start a project | Ask about a retainer |
Notes
What we have learned building AI features, written down for teams who are about to build their own.
Questions we hear
Do you only build AI features? Mostly. We also build the software around them: the API, the admin screen, the dashboard and the deployment. We do not take projects with no AI part at all, because that is not where we are best.
Which models do you use? The one that fits the job, its cost and your rules. We work with models from OpenAI and Anthropic, through Azure and AWS Bedrock, and with open models you can host yourself. We test more than one before we choose.
Where does our data go? Where you decide. We can keep processing inside the EU, use providers that do not train on your data, and sign a data processing agreement. We write down every place your data goes before we build.
Who owns the code? You do. The code, the prompts, the test sets and the documentation are yours, in your repository, from the first day.
Can you work with our own developers? Yes, and we like to. We can build the AI part while your team builds the rest, or we can pair with your developers so they can run the feature without us.
How long until we see something? A working first version usually takes about four weeks. You see the first working round, with its test results, before that.
What does it cost? It depends on the scope. After the first call we send a written proposal with the price, so you know the number before any work starts. The first call is free.
What if AI is the wrong tool? Then we say so. Some problems are better solved with a search box, a rule or a form, and we will tell you before you pay for anything else.