
Sports Analytics Computer Vision System
Master's thesis: a computer vision pipeline that automates two off-the-ball football statistics, body orientation at the pass and scanning checks before receiving.
A collection of research projects, side experiments, and open-source contributions.

Master's thesis: a computer vision pipeline that automates two off-the-ball football statistics, body orientation at the pass and scanning checks before receiving.
AI-powered patient-facing avatar for a hospital setting. Answers patient questions to take routine load off nursing staff, built on Microsoft Foundry agents with a React frontend.
Full-stack work on a global research platform that lets researchers across multiple countries publish, manage, and collaborate on health research projects.

Deep learning pipeline for 3D thoracic organ segmentation from CT scans. nnU-Net 3D vs a custom U-Net variant vs ENet baseline on the SegTHOR challenge.
Predicting Acute Kidney Injury stages in ICU patients with septic shock using Decision Trees, XGBoost, and SHAP explainability on the AmsterdamUMC database.

Research assistant on a 4-session HRI study at VU Amsterdam: a NAO robot assigned each child a personally-fitting book and ran personalized, LLM-supported book discussions across two primary schools. Published at HRI '26.

Research-assistant work on the Dream Robot at VU Amsterdam: an AlphaMini-based social robot designed to deliver guided medical hypnosis to children during invasive procedures in pediatric hospitals. Submitted to IDC '26.
Group project at UvA for Expertwired: automating their manual expert-matching workflow with BERT keyword extraction, BM25 industry routing, and a custom company database.

Optimizing Robobo robot behavior for obstacle avoidance, item foraging, and item retrieval using CMA-ES evolutionary strategies in simulation and hardware.

Computer vision system for Your Neon: converts customer sketches into neon-like visualizations and estimates the LED-strip price. Replaced hours of manual design work per request.

Bachelor's thesis at UvA, built for Saivvy: drone-captured bird's-eye-view dataset, segmentation and detection pipeline, and a tracker bridging both into Saivvy's downstream mapping model.

Personal project: a local AI chat command center on top of LM Studio. Custom agent personas, PDF knowledge bases, web search, all running on my own hardware.
Research and backtesting tooling: a modular ML pipeline for stock prediction with proper temporal splitting, SHAP explainability, and vectorbt portfolio simulation.
Event-driven trading platform: real-time market data via WebSocket, NATS JetStream message bus, microservices architecture for live execution. Sister project to the ML research pipeline.
Classifying 8 gym exercises from smartphone sensor data. Caught a data leak that was inflating accuracy to 100%, then learned what the upper bound actually is.
Two NLP coursework assignments at VU Amsterdam (M.Sc. AI): classical n-gram language models on the Brown corpus, and a feed-forward neural transition-based dependency parser in PyTorch.
Comparing Island Model evolutionary strategies with a constraint-based fitness-sharing variant for evolving neural-network controllers in EvoMan.
Statistical Language and Data Processing project at UvA: testing whether knowledge distillation from GPT-2 to DistilGPT-2 amplifies gender bias when predicting occupation-gender associations.