- Which model gets it right in Dutch?
- Is the best model also the best for you?
- What does one correct answer really cost?
- Can a child talk to it safely?
AI research: what AI can really do.
We benchmark models and study how AI can help with societal problems. Then we put what we learn to use.
See our researchIt started with a study in Artificial Intelligence.
Research at OxVillage grows out of the Artificial Intelligence studies of founder Soufiane El Hassnaoui.
We apply that knowledge in practice: testing models the way a researcher would, and asking what AI can mean for society, not only for business.
We benchmark models. There is no single best one.
Which model is best depends on your task. So we test models on what matters in practice, instead of relying on the numbers their makers publish.
- Real tasks
- Tests based on work people actually do, not only standard lists of questions.
- Dutch
- How well a model handles Dutch language, context and tone.
- Speed and cost
- What one correct answer takes in time and money, not only the price per token.
- Reliability
- Whether a model gives a good answer every time, and how it fails when it does not.
What matters most for your task?
Best for Accuracy: Model A
Illustration only: these are not real results and not real models.
How we test
Start from a real task
What should the model do, for whom, and what counts as a good answer?
Build a fair test
A set of realistic examples, with the answers and criteria fixed before any model runs.
Run every model the same way
Same input, same settings, more than once, so differences come from the model and not from chance.
Look at the mistakes
A score is only the start. We read where models go wrong, because that decides whether you can use them.
How can AI help with societal problems?
That is a research question, not a sales pitch. Whether AI helps depends on the problem, the people and how it is used. These are the principles we work by.
- The problem comes first
- We start with what is going wrong for people, and only then ask whether AI can help. Sometimes it cannot.
- With the people involved
- We work with the people who know the problem, such as teachers and pedagogues, instead of designing from a distance.
- Privacy by design
- What is not needed is not shown. Especially when the people involved are vulnerable or young.
- A person stays responsible
- AI can find patterns and support people, but people interpret them and decide.
Case study: ClassBuddy
Can AI help a school hear what children do not say out loud?
Bullying, worries and trouble at home often stay out of sight. Children do not always tell an adult, and teachers cannot be everywhere. ClassBuddy tests whether a safe conversation with an AI buddy can bring these signals to the surface, without taking away a child's privacy.
Nouri
Friendship
Luma
Anxiety
Rafi
Self-confidence
Zena
Feelings
Appie
Bullying
Chloe
Home situation
Step 1
Children talk
Anonymously, with a buddy, at their own level. No name is attached to what they say.
Step 2
Patterns appear
Teachers see trends across the class, never the conversations.
Step 3
A pedagogue interprets
A person weighs the signals together with conversations and observations in the group.
Step 4
The school gets direction
A written report in plain school language, with concrete next steps.
- 1,500+
- chat sessions in a one-week school pilot
- 6
- AI buddies, each with its own subject
- 5
- domains: group, pupil, team, teacher and parents
- 6 weeks
- from intake to a report with a next step
Got a question worth researching?
A model you want tested, or a societal problem you want to look into with us. We would like to hear about it.