Case studies
Projects we've shipped
Real problems, real data, working systems. No client names (they like their privacy), just what we solved and what changed.
- on-premise LLM · document AICV analysis on an on-premise LLM
- RAG · knowledge graphsLegal answers with citations
- Our project · voice AI · AI agentsA voice assistant on a real phone line
- predictive analytics · machine learningShould we extend this campaign?
- OCR · document classificationDigitising paper employee files
- generative AI · image generationOn-brand campaign images
- Research · RAG · multilingual retrieval · LLM evaluationMeasuring how often a model makes things up
CV analysis on an on-premise LLM
Recruiters couldn't use public AI tools on candidate data. So we built one that never leaves their servers.
- Industry
- HR · 2025
- Area
- on-premise LLM, document AI
The problem
A recruitment team spent hours reading CVs and formatting reports for clients. Public AI tools were off the table: candidate data is confidential, and “we pasted it into a chatbot” is not a sentence anyone wants to say in an audit.
What we did
An application on a local language model that does what the experts used to do by hand: it pulls the key information out of CVs, prepares preliminary candidate recommendations and writes short summaries. It runs on the client’s own infrastructure, so the data never leaves the company.
Results
- Several teams now review and sign off on summaries instead of reading every CV and retyping it into a report.
Legal answers with citations
A search engine for legal questions that shows the exact provisions behind every answer.
- Industry
- Legal · 2025
- Area
- RAG, knowledge graphs
The problem
Finding precise information in a maze of regulations took time and expertise. And in law, a chatbot that only sounds right is worse than no chatbot at all.
What we did
A RAG chatbot: it first finds the right provisions and hands them to the model together with the question (only a bit of magic, mostly math). The provisions come from curated knowledge graphs of the regulations, a map of which rule connects to which. So every answer is grounded in the actual acts, and the system checks whether the organization meets the requirement.
Results
- A legal Q&A engine that cites the acts behind every answer, so each answer can be checked against its source.
A voice assistant on a real phone line
Our own project, built to show how we connect AI with technology that already exists. Here, real telephony.
- Industry
- Telecom
- Area
- voice AI, AI agents
Why we built it
Most AI projects live or die on integration, not on the model. So we built a voice agent that has to work with something unforgiving: a real phone line.
What it does
It answers calls, talks to the caller and gives information. When a person should take over, it sends a summary of the conversation by e-mail so someone can call back.
What’s next
We plan to put its number on this site, so you can call it the same way you chat with the assistant on our home page.
Results
- A working voice agent on a real phone number, which we plan to make available on this site.
Should we extend this campaign?
A model that tells which marketing campaigns are worth extending, before more budget goes in.
- Industry
- Marketing · 2025
- Area
- predictive analytics, machine learning
The problem
Deciding whether to extend a campaign was a mix of experience, spreadsheets and gut feeling.
What we did
A predictive model that analyses how a campaign affects click-through rates and shows which campaigns are worth extending.
Results
- Better campaign returns, without anyone tuning parameters by hand.
Digitising paper employee files
Boxes of paper HR documentation turned into a searchable, structured archive.
- Industry
- HR · 2025
- Area
- OCR, document classification
The problem
A large amount of paper employee documentation had to become digital, and usable, not just scanned into a folder nobody opens.
What we did
Text extraction from the scans (OCR) combined with a model that recognises the type of each document, so every file lands in the right place with its key data pulled out.
Results
- An organised digital collection of employee documents with structured data, instead of shelves of binders.
On-brand campaign images
An image generator that sticks to the brand book, so designers don't redo the same banner fifty times.
- Industry
- Marketing · 2025
- Area
- generative AI, image generation
The problem
Campaigns needed many images in many variants, and the brand book left little room for creative accidents.
What we did
A generator of realistic graphics that strictly follows the client’s brand book.
Results
- Designers freed from repetitive production work and back to the creative part.
Measuring how often a model makes things up
A year-long, EU-funded research project in three languages (English, Polish, German): how to teach a retriever a new domain, and how to measure how often a language model makes things up.
- Industry
- Research · EU-funded · 2025
- Area
- RAG, multilingual retrieval, LLM evaluation
The question
RAG is only as good as what it finds. Off-the-shelf retrievers get lost in specialist domains, and outside English they get lost faster. There’s rarely labeled data to fix that. And without a solid metric, “it hallucinates less” is just a feeling.
What we did
We trained retrievers in English, Polish and German with GPL (generative pseudo labeling): the model writes its own practice questions for unlabeled documents and learns the domain from them, so nobody has to label thousands of examples. We built the training tools and datasets, evaluated the models, measured their hallucination levels and researched the metrics themselves. Along the way we worked on controllable text generation.
Results
- Tools and datasets for training domain-specific retrievers in three languages, and a way to measure hallucinations instead of guessing. The same thinking goes into every RAG system we build for clients.
Similar problem on your side?
Tell us about it. We'll reply with an honest assessment and a first step.