Machine learning development

Machine learning is a form of AI in which a model learns patterns from examples in data, rather than from rules written by a person. It uses what it has learnt to predict or recognise something in new cases.

In supervised learning, the model learns from examples with known outcomes: a category (classification) or a number (regression). In unsupervised learning, there is no known outcome, and the model finds groups or anomalies itself. Machine learning is one of the AI applications we build.

Machine learning is a family of techniques, and we have a separate page for each type. You can count or inspect objects in images with computer vision; for GVB, it is used to count passengers in footage from security cameras. Categorising or searching text falls under NLP and text analysis.

Predicting future demand or stock levels is called forecasting. With predictive maintenance, you use asset data to predict when maintenance is needed. For planning optimisation, data science and optimisation help create a workable schedule. Finding anomalies, such as an incorrect invoice, is called anomaly detection.

By comparing it with a simple baseline, such as always predicting the most common outcome or using the rule you currently follow. If the model does not clearly outperform that baseline, it is not worth using. For time-ordered data, we train on the past and test on a later period; otherwise, the model learns from the future.

When an outcome is rare, the proportion of correct answers tells you little. That is why we measure precision — the proportion of alerts that are justified — and recall — the proportion of actual cases the model finds. Which matters more depends on the cost to you of a missed case or a false alarm.

Data leakage means the model uses information during training that would not yet be available when making a prediction. It then scores too well in tests and performs poorly in production. One example is a field that is only filled in after an invoice has been rejected.

That is why we split the data into training and test sets before processing it, and check when each column becomes available. The same preprocessing runs in production as during training; otherwise, the model receives different input from what it learnt on.

A model learns from historical data. When a new product, a different way of working or a different customer group comes along, the input changes and predictions become less accurate. This is called drift. That is why we monitor inputs and predictions, and define when the model should be retrained.

Before going live, also establish who owns the model, who is alerted to anomalies and who approves a new version. You can read about the technical side under MLOps. Your team can handle management, or we can take it on through management and continuous improvement.

We build where your data already resides. On Databricks, you can track experiments and models with MLflow and manage model versions in Unity Catalog. In Microsoft Fabric, you can track training runs with its built-in MLflow and write predictions to OneLake, where Power BI can read them. With LUMC, we work on on-premises, sovereign infrastructure.

This keeps your options open: MLflow is open source, and the code and models remain in your own version control and environment. As a Databricks and Microsoft partner, we set up the data and AI platform so your team can take it over.

If the model processes personal data, the GDPR applies. According to the Autoriteit Persoonsgegevens (AP), a data protection impact assessment (DPIA) is mandatory when you use algorithmic systems involving personal data, even in a pilot. People have a right to human review: you cannot make a decision that significantly affects them solely through automated processing. For public authorities, the AP points to the Algoritmeregister.

The EU AI Act classifies software that sorts CVs and systems that calculate credit scores, among others, as high-risk. According to the European Commission, strict obligations for these systems will apply from 2 December 2027. Your legal or compliance team must assess whether your model falls into this category. We help prepare the inventory and risk classification; read more under AI Act.

When a fixed rule already makes the decision well enough; when the outcome is not recorded anywhere, leaving the model nothing to learn from; or when the decision occurs so rarely that the cost of building and managing a model outweighs its benefits. In those cases, a rule in the existing system or a good dashboard makes more sense.

A model will not fix a messy data source either; start with the data foundation instead. If you do not know whether the problem lies in the data, processes or organisation, start with a maturity assessment. If a simpler solution fits, we will say so.

Want to compare companies first? See our overview of the best machine learning companies.