One partner. Who does the building.
Martin Nørmark Hansen, founder of algoritma. A peer-reviewed researcher in how language models and machine-learning algorithms learn, and one of the few who both publishes AI research and puts the models into production for Danish businesses. About as close to a real AI specialist as it gets: someone who understands the theory in depth and builds it all the way, from the first data pull to a model in production.
Models in production. Not prototypes in a drawer.
A selection of the deliveries I have built for Danish businesses. I do not name clients here, that is agreed confidentiality, but I am happy to walk through the methods, data and results in detail on a video call.
found in three years of order history using unsupervised clustering (K-means and HDBSCAN on RFM features, validated with silhouette scores). One segment accounted for over 40% of revenue, and once marketing began targeting it, repeat purchases rose measurably within the following quarter.
Clustering · order history · in productionlower forecast error (MAPE) than the spreadsheet purchasing used to steer by. Gradient boosting with calendar, campaign and seasonality features, validated with rolling backtests. Per-SKU forecasts six weeks ahead mean fewer stockouts and less capital tied up in inventory.
Time series · ERP data · in productionof the fields on incoming documents are now read automatically by a fine-tuned OCR pipeline with document classification and field extraction, with confidence thresholds so edge cases always reach a human. What used to be a standing data-entry task is now fifteen minutes of spot checks a week.
OCR · classification · in productionSpecialised in machine learning
MSc in data science from SDU, specialising in AI and machine learning: deep, technical training in how models actually learn, from neural networks and gradient boosting to clustering and language models. Not just how to use the tools, but how they work under the hood. On top of a business degree in project management from CBS.
That combination is what makes algoritma rare: one person who researches how AI works at the deep level, builds the model, and runs the project that gets it into production. No handover between adviser and developer, nothing lost in translation.
Research into how models learn, and how people come to trust them.
I keep one foot in research and one in production. Two strands of research sit behind the way algoritma delivers.
Publication · accepted at BlackboxNLP 2026Subspace Clustering and the Representational Utility of Quanta Fingerprints
Peer-reviewed research on how best to uncover the small, discrete skills ("quanta") a language model learns, by comparing four clustering paradigms. The finding: the method that best fits the theory is not the one that yields the most useful features. The code is publicly available.
Read the paper (PDF) →What language models actually learn
Research into how a language model's capabilities can be broken down into small, distinct skills by clustering the model's per-token gradients. I compared four clustering paradigms (spectral, sparse subspace clustering with Lasso and OMP, and hierarchical) across two transformer sizes and showed that the method that best fits the theory is not the one that yields the most useful features. The same cluster structure served as a "fingerprint" that, in a controlled experiment, distinguished AI-generated text from human writing almost without error. The code is publicly available.
Trust in AI within organisations
A field study of an international GenAI implementation: how do you build trust in a black-box model among the people who have to use it? The finding was clear. Trust does not come from technical explanations alone, but from formal controls around the model's output, combined with a key person who translates between technology and business. That principle is built directly into how algoritma delivers.
Let's have a chat.
You will get concrete suggestions, whether or not we work together afterwards.