RESEARCH
Research
Can an LLM agent find the privacy risks in synthetic health data that standard tests rate as safe?
MSc thesis · in progress
Agentic Privacy Auditing of Synthetic Health Data: Evaluating LLM Audit Agents against Attack-Based Re-identification Metrics
Synthetic health data is meant to keep the patterns of real patient records without containing any real patient. Before it is shared, someone has to check that no synthetic record still points to a real person.
A common quick check measures how far each synthetic record is from the closest real one, which is what the figure shows. Stronger checks run real attacks against the data and count how often a person can be re-identified.
My thesis asks whether an LLM agent that audits the same data can find risks that those attack-based metrics rate as low. The work is in progress and there are no results to report yet.
Fig. 1 · Distance to the closest record
Each dot is a synthetic record. Each ring is a real person's record. A dot that lands on a ring can give that person away. Finding those cases is what my thesis is about.
- Real record
- Synthetic record
- Too close: 7 of 92
Move the pointer to audit a record. Illustrative data, not study results.
Work so far
Research and academic work
- In progress
MSc thesis
Agentic Privacy Auditing of Synthetic Health Data: Evaluating LLM Audit Agents against Attack-Based Re-identification Metrics
- Done
Enterprise AI agents
Compared four 2026 surveys where reported adoption ranged from 17% to 65%, and traced why.
- Done
Agent design patterns
ReAct, Reflection, Planning, Tool use and Multi-agent, matched to real deployments and to MCP and A2A.
Interests
Where my curiosity goes
- Healthcare AI and privacy
- Software that is useful to a clinic without moving patient data. Klindok was my way of learning what that takes.
- Evaluating LLMs and agents
- Testing what a model or an agent really does before relying on it, and being clear about what was not tested.
- Robotics
- What changes when a model runs on a real device. I am new to this field and learning it through Kiko.
- Digital twins
- Simulated copies of physical systems, for testing behaviour before the hardware exists. An interest I am starting to explore.
Paper breakdowns
Papers I have read closely
Published whenever I finish reading a paper properly. No fixed schedule.
- arXiv preprintMust read
The DCR Delusion: Measuring the Privacy Risk of Synthetic Data
The most popular privacy test for synthetic data passes datasets that membership inference attacks show are leaking.
Open to research collaborations
LinkedIn