2020–2025 · TU Delft
My PhD
Towards next-generation meeting support systems
Meetings are one of the most information-dense interactions humans have — and one of the most under-supported. We remember only a fraction of what was said, important decisions get lost, and follow-through is often patchy. I wanted to understand whether AI could help: not by replacing human judgment, but by augmenting it.
My PhD asked: can a system learn to identify what is important in a meeting from your verbal and non-verbal signals?
The dataset
I spent the first year building a 700 GB multi-modal time-series dataset capturing meeting interactions across speech, language, gesture, gaze, and physiological signals. Real teams, real tasks, real pressure.
Collecting, annotating, and cleaning that data taught me more about the practical realities of ML research than any textbook could — annotation pipelines, inter-rater reliability, and all the edge cases in between.
The models
I built and evaluated ML pipelines to predict subjective importance ratings from multi-modal signal streams. The systems achieved above-chance performance across more than 10,000 simulation runs — a meaningful result given the inherent subjectivity of importance as a construct.
What I learned
A PhD teaches you to be comfortable with uncertainty. Most experiments don't work the first time. Most papers get rejected at least once. The work is slow, iterative, and often invisible to the outside world.
It also taught me to think rigorously about ethics — especially when AI systems observe and interpret human behaviour. Consent, privacy, model bias, and the gap between benchmark performance and real-world deployment became central to how I approach any AI project.
Recognition
- Outstanding Early Career Scientist Paper — IEEE Ro-Man, 2022
- TAILOR Connectivity Grant — University of Cambridge, 2023