80% direct response from prize winners through AI-driven selection of thousands of campaign entries
Harman
Data science consultancy in Haarlem
We build models that predict and assess, testing them against your employees' judgement, as we did with the submission selection process we developed for Harman.
For Harman, we developed a selection process that assesses entries objectively and reproducibly
Manually assessing entries for international prize campaigns took Harman a lot of time. Our data-driven selection process uses GenAI and LLM technology to assess large numbers of entries consistently. Harman is not based in Haarlem; the case study shows how we build an assessment model.
How do we assess every application, case file or response consistently?
At local authorities, healthcare organisations and business service providers in Haarlem and Zuid-Kennemerland, staff read through stacks of applications, case files and open-ended responses. That takes time, and two colleagues sometimes assess the same case differently. A model makes a suggestion; your staff member decides.
Assess, predict, match and monitor: models that support your team’s work
Assess against set criteria
For Harman, our selection process assesses submissions objectively and reproducibly, so large numbers are selected in the same way.
Predict an outcome
As part of the St. Antonius AI team, we developed an algorithm that supports doctors in deciding whether a patient should be admitted to hospital.
Match and classify
For recruitment agency Olympia, we used a classification model to make matching candidates with suitable job openings more efficient and accurate.
Monitor after deployment
For the AI models Rijkswaterstaat has in production, we developed a generic monitoring package to detect drift and outliers.
Your employees’ judgement is the benchmark we use to test every model
-
Decision and criteria
Which decision the model supports, what criteria your employees currently use to make their assessments, and who uses the output.
-
Testing against human assessments
We compare the model’s output with previous assessments and ask your employees to review the differences.
-
In use and monitored
The model is integrated into the process, with an employee making the decision, and we monitor whether it continues to produce accurate results.
Is our assessment work suited to a model, or would a fixed rule suffice?
A good fit if
Your team repeatedly assesses or predicts similar cases, previous decisions and their outcomes have been recorded, and someone genuinely wants to use the result in their work.
Not a good fit if
There are few cases, every case is unique, or the criteria are clear enough for a fixed rule in your system to do the job. In that case, we advise against using a model.
What determines the scope
How many types of cases there are, whether previous assessments have been recorded in a usable way, where the result needs to appear, what privacy requirements apply, and who will manage the model afterwards.
Decisions about citizens or clients?
A staff member remains responsible for the decision, and we document what the model is based on. Your data protection officer determines whether a DPIA is needed.
A decision rule, a language model or your own trained model: which fits your assessments?
A decision rule
Choose this if the criteria are fixed and there are few exceptions. A rule is easy to explain and maintain.
A language model
Choose this for free text, such as submissions or open-ended answers, when there are few labelled examples. You still need to test it against previous assessments.
Your own trained model
Choose this if you have many previous decisions with recorded outcomes. We recommend it if you need to show which features carry the most weight.
From Westerdoksdijk in Amsterdam, we also work with organisations in Haarlem and Kennemerland
Data Science Lab has worked with data and AI since 2016 from our office in Amsterdam, close to Haarlem. We’ve built models for St. Antonius, Rijkswaterstaat and Harman, among others, and share knowledge at our Techdays.
The decision your employees currently make manually is the focus of our first conversation
Stef Stoelwinder
Business DevelopmentStef is your first point of contact. In our first conversation, we discuss which decision needs to be made better or faster, what previous assessments are available and whether a model is a good fit.
E-mail info@datasciencelab.nl
Phone +31 (0)20 3636 724
Which assessment takes up the most of your team’s time right now?
Tell us which cases your team assesses and what data is involved. We’ll reply by email or arrange an initial conversation.
Message sent
Thanks for your message. We’ll be in touch as soon as possible.
Frequently asked questions
It builds models that use your own data to predict, assess or categorise cases, then puts those models to work. In Haarlem and the surrounding area, this often involves municipalities and the province, healthcare organisations and business service providers handling many similar cases.
If you’re looking for data science in Haarlem or data science consultancy in the region, we provide the team to build, test and hand over the model. You can find all our applications under AI applications.
No. We’re based on Westerdoksdijk in Amsterdam. For organisations in Haarlem, Heemstede, Bloemendaal and the rest of Kennemerland, we can visit you or work online.
For AI that we build and then manage, read about our AI consultancy in Amsterdam.
We take past cases with known decisions and ask the model to assess them again without knowing those decisions. Where the model and the employee disagree, your staff review the case: sometimes the model is wrong; sometimes two colleagues turn out to disagree.
Only once we understand the difference do we use the model to support the work.
A language model suits unstructured text, such as submissions, open-ended answers or letters, especially when there are few labelled examples. For Harman, we worked with GenAI and LLM technology.
If you have structured data with many recorded outcomes, a trained model is often easier to explain and cheaper to use. Read more under machine learning.
We include only the data needed for the purpose and document which features the model uses. An employee remains responsible for decisions about citizens, clients or candidates.
Your data protection officer assesses whether the processing is permitted and whether a DPIA is needed. What we provide to support that assessment is set out under DPIA for AI applications.
A model can quietly become less effective as cases change, for example because of a new policy or a different target group. That’s why we monitor inputs and predictions and agree when to retrain the model.
For Rijkswaterstaat, we developed a generic monitoring package that checks all models in production for drift.
Yes. In Utrecht, the focus is on predictive models in healthcare and service delivery; see data science consultancy in Utrecht. In The Hague, it’s on public organisations and healthcare; see data science consultancy in The Hague.
Our approach is the same everywhere: first the decision, then the data, then an honest assessment.
When there are too few cases to learn from, when the criteria can be captured in a fixed rule, or when no one will use the outcome in their work. A model also needs maintaining.
We’ll tell you that, too. Sometimes better record-keeping or an overview of past assessments is the first step.