Theses, personal projects and course projects. Each page describes the problem, the approach and the results.
A public job board that reads postings from Dutch employer career sites, classifies tech roles and extracts fields such as skills, seniority and visa sponsorship. I designed the system and maintain it in production, with an evaluation gate in CI and a Kubernetes deployment on AWS.
A browser demo that tracks the hand with MediaPipe and renders a particle effect for each of five gestures. I implemented the gesture recognition and the effects, which run in real time in plain JavaScript.
Measures the social inclusion of preschool children with hearing loss from wearable positioning and audio data. I fused the sensor streams, identified six activity contexts with a Gaussian Mixture Model and compared the two groups of children within each context using mixed-effects models.
Applies synaptic tagging and capture, a theory of how memories become permanent, to networks trained with backpropagation. I designed the mechanism and evaluated its consolidation log as an explanation of what the network retains.
A group project presenting an interactive visualisation of earthquakes and tsunamis, with a linked map, timeline, scatterplot and detail view in which a selection in one view filters the others. I developed the timeline, a brushable heatmap with drill-down navigation.
A group design of a smart home system that learns household routines, including a proof of concept for loss-free sensor data ingestion during network outages. I developed most of the proof of concept, a FastAPI service and RabbitMQ queue fed by buffering sensors.
In a group of four, we reimplemented an underwater image enhancement network from its paper, as no code was publicly available, and reported which results did and did not replicate. I implemented the model architecture in PyTorch and wrote the blog post.
A group study of whether three LLMs labelling meeting transcripts agree more when they see each other's earlier annotations. Agreement rose from 0.185 to 0.216. I implemented the experiment loop that runs the panel and wrote most of the prompts.
Classifies nettle-family pollen from microscope Z-stacks using every focal plane. I trained 3D CNNs in PyTorch, and a ResNet3D trained on the full stack reached 95% test accuracy.
Reads Dutch bank CSV exports, identifies the recipient of each payment and detects recurring payments. I developed the merchant matching and recurring-payment detection, which reach a B-cubed F1 of 0.95 on held-out data.