Skip to content

AI agency in Den Bosch

We build AI that assesses large numbers of requests or submissions consistently, based on your criteria. That’s how Harman selects submissions using an LLM.

For Harman, we developed a GenAI-based selection process that assesses entries objectively and reproducibly

Large numbers of entries for international prize campaigns are now selected consistently. The marketing team saves time and can scale up its campaigns.

Read the case study

Do two colleagues really assess the same request in the same way?

For many requests, submissions or reports, the outcome depends on who assesses them and when. A model that applies your criteria takes the same approach every time and explains why.

Schedule a conversation

Your criteria, applied consistently every time

Discuss your assessment

Define the criteria

Document what matters and how much weight it carries with the people who currently assess submissions, including examples of good and bad.

A score with an explanation

A score and a reason for each application or submission, so an assessor can see what the model bases its score on.

A person makes the decision

The model ranks and flags submissions, a member of staff makes the decision, and uncertain cases always go to a person.

Repeatable at scale

Just as Harman has thousands of submissions assessed in the same way, we test the model against a fixed set.

First see how you assess cases now, then test the model against that

  1. Criteria and case set

    We work with your assessors to define the criteria and select past cases with known outcomes.

  2. Model alongside human assessors

    The model assesses the same cases, and we compare where it differs from the human assessors and why.

  3. Putting it into use

    First as guidance for the assessor, then at scale, with regular checks on samples.

Can a model help decide on the applications we receive?

Put it to the test in a conversation

A good fit if

You receive many text-based submissions, applications or reports, the criteria can be defined, and reviewing them currently takes a lot of time or varies from person to person.

Not a good fit if

There are only a handful of cases each year, or every decision calls for a judgement that cannot be captured in criteria. Then people need to make the decisions.

What determines the scope

The number of cases, how many criteria there are, which systems supply the applications and how thorough the subsequent review needs to be.

Is it fair and explainable?

That is what we assess: we compare results with previous decisions, look at groups that receive different outcomes and record the reason for each judgement.

Reviewers with a checklist, fixed rules or a language model with criteria?

Reviewers with a checklist

Choose this when there are few cases and specialist knowledge is decisive. You accept that assessments will vary between reviewers.

Fixed rules in a system

Choose this when the criteria are measurable, such as an amount or a date. Rules fall short when dealing with free text and justifications.

A language model with criteria

Choose this when many applications consist of text. Our advice: start by using it to advise your reviewers and compare the results.

We've built generative AI that runs within business processes before

Data Science Lab has been around since 2016 and is a partner of Microsoft and Databricks. We've built LLM applications for Harman and Fokker, and share what we learn at our Techdays.

Book an introductory meeting

We’ll discuss the assessment that takes your team the most time first

Stef Stoelwinder

Stef Stoelwinder

Business Development

Stef is your point of contact. In our first conversation, we’ll discuss what your team assesses and which criteria you use.

Applications, submissions or reports: where would you like to start?

Tell us what comes in and how your team currently assesses it. We’ll reply by email or arrange an initial conversation.

Enter a valid email address.

Let us know how we can help you.

Frequently asked questions

It builds AI applications that run within your own processes and manages them. Here, the focus is on assessing and selecting applications, submissions or reports against set criteria.

You can read about more applications at AI applications.

The model receives your criteria and examples, reads each application and gives an assessment for each criterion, with a reason. We compare the results with previous decisions made by your assessors.

Read more about generative AI.

We design for consistency using fixed instructions, set criteria and a test set that runs again after every change. That is how Harman selects large numbers of submissions in the same way.

If the model deviates, we see it in the test.

That depends on what is being assessed. If it involves people, such as job applicants or benefit claimants, stricter rules apply. Read about them on our AI Act page.

We discuss this in advance, before anything is built.

Data Science Lab is based in Amsterdam. We visit for the initial work and to work with your assessors; we do the rest remotely.

We prefer to sit down in person with the people who currently do the assessing.

It runs in your own cloud environment or a restricted-access service, with agreements on what is retained. If personal data is involved, we identify the risks in advance.

Applications are not used to train a public model.

Then our data science agency for Den Bosch is a better fit: it builds models that calculate in advance how a decision will affect your planning.

The two approaches can also be combined.

Start with our AI consultant for Den Bosch. Together, we identify the question with the most value, then we build.

We also agree in advance who will manage what.