A generic monitoring package for machine learning models developed for Rijkswaterstaat
Rijkswaterstaat
AI consultant in Utrecht
AI advice for rail, roads, water and healthcare, through to monitoring models in production, as we do for Rijkswaterstaat.
One package to monitor all Rijkswaterstaat models in production
Rijkswaterstaat has several AI models in production. Monitoring them manually does not scale, so we built a single package for drift and outlier detection. “DSL has come up with an inventive solution that we can put to good use.” Nanine Overgaauw, SRE Data Science, Rijkswaterstaat
Who notices when a model in the network gradually drifts?
ProRail and NS are headquartered in Utrecht, and Rijkswaterstaat also operates from Westraven there. A model used in the rail or road network has to contend with new equipment, replaced sensors and work along the route. Without monitoring and someone responsible for the model, its quality can decline unnoticed.
Models for rail, roads, water and healthcare must remain accurate for years
Use cases from your own sensors
Fibre-optic cables, track inspection trains, cameras and fault reports already generate data. In ProRail’s Datalab, we explored what machine learning could do with fibre-optic signals.
Monitoring that evolves with the models
When developing a new model, we plan checks for drift and outliers from the outset. If models are already running, we can add those checks later, as we did for Rijkswaterstaat.
Models that work alongside doctors
As part of the AI team at St. Antonius Ziekenhuis, we built an algorithm to support emergency department doctors in deciding whether to send a patient home or admit them.
Deciding on management during development
For each model, we agree during development who will monitor it, who will retrain it and whether your own team will take it over later.
The stretch of track, the road or A&E comes before the model
-
Choose one decision
For example, when a set of points needs maintenance or which A&E patient needs a bed—and who makes that decision now.
-
Check data from the field
Sensor, inspection or patient data: is there historical data, can it be used, and what happens if the sensors or working methods change?
-
Build with ongoing management in mind
We build the model with monitoring and a retraining schedule, and document who approves a new version.
Does your network or hospital need an AI consultant right now?
A good fit if
You manage rail, roads, water infrastructure or healthcare processes, have sensor, inspection or patient data, and are running a pilot or model without a clear plan for who will maintain it after handover.
Not a fit if
A fixed threshold in the maintenance or reporting system already solves the problem. Or the model is running and only monitoring is missing: in that case, start with MLOps, without a consultancy project.
What determines the scope
The number of models and data sources, whether you already have a platform and monitoring in place, information security and procurement requirements, and how much of the management you want to handle yourselves.
You can take over later
Where possible, we build in your existing environment and provide documentation. If your team wants to take over management later, we agree in advance what will be handed over.
Who manages the model after delivery?
Your own team
Choose this if your data team has experience with models in production, the model runs on your own platform and someone regularly handles drift alerts.
Data Science Lab
Choose this if we build the model and you do not want to set up a management team. We only manage models we have developed ourselves, through AI managed services.
Work together first, then hand over
Choose this if your team wants to take over management but has not yet had a model in production. Our advice: set up MLOps together, then hand it over.
We stay involved when the model is running in production
Data Science Lab has worked on data and AI from Amsterdam since 2016, as a Databricks and Microsoft partner. We plan for ongoing management during development and share knowledge in our lab and at Techdays.
We discuss ongoing management in the first conversation
Stef Stoelwinder
Business DevelopmentStef is your first point of contact. In our first conversation, we look at the model or idea, the data and who will manage it going forward. We can meet at your offices, in Amsterdam or online.
E-mail info@datasciencelab.nl
Phone +31 (0)20 3636 724
Is a model already running, or is it still just an idea?
Describe the model, pilot or idea and who currently manages it. We’ll reply by email or suggest an initial conversation.
Message sent
Thanks for your message. We’ll be in touch as soon as possible.
Frequently asked questions
Several infrastructure operators and hospitals in the region have a team like this. ProRail has a Datalab, Rijkswaterstaat brought its digitalisation, data and AI units together in a Data and AI Lab at the end of 2025, and UMC Utrecht works with AI labs and the Health Data Space Utrecht.
We work alongside that team, as we do with ProRail’s Datalab, or as part of it, as we do in the AI team at St. Antonius Ziekenhuis. Our office is in Amsterdam; we work at your premises or online.
Fiber-Optic Acoustic Sensing (FOAS) uses fibre-optic cables to measure sound waves. Together with ProRail’s Datalab, we used machine learning, including Databricks and Tensorflow, to automate the detection of trains, track components and wheel defects. The aim: a safer, more sustainable rail network.
An initial investigation shows whether a signal is useful. Using it in the maintenance process raises further questions: who responds to an alert, how often the model is retrained, and who monitors it when cables or rolling stock change. Read the ProRail case study.
Drift means that a model’s inputs (data drift) or predictions (prediction drift) gradually shift away from its training data. In a network, this can happen because of new equipment, a replaced sensor, major maintenance or a change of season. The model still produces results, but they become less accurate.
Rijkswaterstaat has several models in production, and monitoring them by hand does not scale. The package we built is designed to check them all for drift and outliers, using Python and Alibi Detect, among other tools. To see how this fits into a model’s lifecycle, read about MLOps.
Which model version is running and which training run it came from; the thresholds for drift and data quality; who receives an alert; how retraining works; and who decides whether a new version can go live. Without these agreements, a handed-over model remains dependent on one person’s knowledge.
According to Rijkswaterstaat’s AI strategy, human oversight and judgement remain essential. A handover should therefore also specify who reviews a result before anyone acts on it. See also data and AI for government.
Above all, a clear decision point: when does an emergency doctor know enough to anticipate an admission? For St. Antonius Ziekenhuis, which has emergency departments in Utrecht and Nieuwegein, we developed an algorithm as part of its AI team. It can help doctors decide whether a patient can go home or needs to be admitted.
The aim is to inform hospital departments earlier about a possible admission, so they can prepare sooner and the patient can move from the emergency department to the ward more quickly. “Professional collaboration with committed specialists,” says Marleen Vreeburg, an emergency doctor at St. Antonius. Read the case study or learn more about data and AI in healthcare.
It may be. Annex III of the AI Act lists AI systems intended as safety components in the management or operation of road traffic or in the supply of water. For railway systems and medical devices, the relevant provisions are in the product legislation listed in Annex I. Whether your model falls within scope depends on what it does.
Under the AI omnibus of July 2026, the Annex III requirements will not apply until 2 December 2027. We can help you take stock and prepare a risk classification; your legal counsel or compliance department determines the legal interpretation. Read more about the AI Act.
When a fixed rule already solves the problem: a threshold in the maintenance system, an alert at a particular measurement or a useful dashboard built on inspection data. A model only adds value when the relationships are too complex for such a rule, or there are too many signals to assess manually.
If the model is already running and only monitoring is missing, start with MLOps. If what you mainly need is extra capacity in your own data lab, bringing in an embedded specialist such as an MLOps engineer or data scientist makes more sense than a consultancy project.