Thesis project: On-premise AI
Fine-tune a small model. Keep it in the building. Many Dutch SMEs would rather keep sensitive data off US cloud AI, but running a large open model on-premise is out of reach for them. Small language models fit on affordable hardware, yet out of the box they underperform on specific tasks. You investigate whether fine-tuning closes that gap.
- Thesis project
- MSc
- 2 places
- Amsterdam
About Datagrove
Datagrove is an AI-native development agency on the Herengracht in Amsterdam. We build AI systems, platforms and apps for businesses, and we keep hearing the same question from clients: how do we use AI without handing our data to a large US provider?
This isn't hypothetical. AI infrastructure and frontier models are dominated by a handful of US companies, China is building its own ecosystem, and Europe risks falling behind both. Competing at frontier scale is unrealistic. Making smaller, open models genuinely deployable is not.
What you do
The research question: to what extent can fine-tuning an open-source small language model on a representative SME agentic task match the accuracy of a much larger, general-purpose open model on that same task, while running fully on-premise on commodity hardware? The proposed method:
- Select one open-source SLM (for example in the 1 to 8B parameter range) and one representative agentic task.
- Fine-tune the model on a task-specific dataset, with a clear evaluation protocol.
- Deploy the fine-tuned model on Datagrove's on-premise hardware and benchmark it against a larger, un-tuned open model on the same task.
- Answer the research question with a quantified result and discuss how the approach generalises to other tasks.
Who we are looking for
This project is listed on the UvA Thesis Matching Platform and is open to master's students of these programmes:
- MSc Artificial Intelligence.
- MSc Computer Science (joint degree).
- MSc Information Studies, Data Science track.
- You have experience with Python and machine learning, and want to actually train and deploy a model, not just describe one.
- Start as soon as possible. 1 January 2027 works too, when most programmes begin their thesis semester.
What we offer
You research a question European businesses are asking right now, with real data and real hardware. Research is not a side activity for us: we want to keep learning ourselves, and this project is part of that.
- A real SME use case and sample data from an existing client engagement.
- On-premise hardware for the fine-tuning and deployment experiments.
- Daily supervision by Sion Coolwijk, UvA alumnus MSc Data Science, with applied AI/ML experience.
- Freedom in how you carry out the research, guided by a daily supervisor. The research question is a starting point, not a fixed frame: if another domain or method proves stronger along the way, there is room to adjust the direction.
- A desk on the Herengracht in Amsterdam.
- Two places, so room to split the research with a fellow student.
Apply for this role
Leave your details, write a short motivation and attach your resume. We read every application ourselves and respond quickly.