AI agency in Utrecht

An AI agency, or AI company, builds applications that use data to support decisions and keeps them running in production. In Utrecht, this often involves networks: ProRail has its headquarters at Moreelsepark, and the province owns the tram system.

We build these models as part of our AI applications. If you first need advice on which application to choose, read about our AI consultant in Utrecht.

With Fiber-Optic Acoustic Sensing, fibre-optic cables measure sound waves. The ProRail case study lists techniques including Fourier transforms, image segmentation using a U-Net model and Tensorflow, on Databricks and Azure. A Fourier transform breaks a signal down into its component frequencies.

For day-to-day operations, other questions arise: how many false alerts are acceptable, who responds, and what happens when a cable is replaced? Read the ProRail case study.

Version control for code and models, a model registry, a pipeline that tests and deploys new versions, and monitoring that sends an alert to whoever is responsible at the time. You also need an agreement on what the operations team does if the model fails.

At Waternet, as part of the Datalab team, we set up and maintain an MLOps workflow with an MLflow model registry and CI/CD templates. Read about how it works under MLOps.

Its input changes. A new cable, a different type of wheel or a renewed section of track produces signals the model does not recognise. It still generates alerts, but they become less accurate. This is called drift.

That is why we monitor inputs and predictions and retrain the model when needed. For GVB, we established an MLOps lifecycle because its segmentation model needed to stay up to date as things changed over the years.

Yes, provided there is a history of maintenance and faults that you can compare with the measurement series. According to the province of Utrecht’s Beheervisie Tramsysteem 2020-2024, the province owns the assets, while the Trambedrijf Beheer en Onderhoud team manages the local rail infrastructure.

With maintenance records for trams, rails and points, you can predict where maintenance is needed. Read more about predictive maintenance and AI for infrastructure management.

Possibly. Point 2 of Annex III of the AI Act covers systems intended as safety components in the management and operation of critical digital infrastructure, road traffic, and the supply of water, gas, heating or electricity. Rail systems fall under the European product rules in Annex I.

Under the AI omnibus, the requirements of Annex III apply from 2 December 2027. Legal interpretation is for your lawyer or compliance department; we can help identify the relevant systems and propose a risk classification. See AI Act.

Yes, if your team has experience with models in production. We then agree in advance who owns the model, who receives alerts, when it is retrained and who approves a new version. If you lack the capacity, you can hire an MLOps engineer.

If we continue to manage it, we do so through AI Managed Services. We only manage solutions we have developed ourselves.

If a fixed threshold in your monitoring system already detects the anomaly reliably, or if the operations team mainly needs a clearer overview, a dashboard built on your existing data will help more than a model.

If you do not yet know whether the signals contain enough information, explore the data first. That is what our data science agency in Utrecht is for. Building a model only makes sense if you can demonstrate a relationship.