A curated set of research, infrastructure, and web projects. Each one explains the problem it solved, the stack it runs on, and what I actually contributed.
Master's thesis at Amsterdam UMC. Deep learning models to reconstruct dynamic MRI scans from undersampled measurements — making scans faster without sacrificing clinical image quality. Full pipeline from raw k-space data to reconstructed image sequences.
Group project for a recommender-systems challenge, focused on ranking news recommendations and comparing model performance against baseline approaches. The project implemented preprocessing, training, and evaluation pipelines, with results reported using AUC, MRR, nDCG@5, and nDCG@10. Experiments compared random ranking, NRMS, GERL, and a weighted neighbour-sampling variant to understand where graph-based recommendation helped — and where extra sampling complexity did not.
Course project on 3D medical image segmentation using the SegTHOR challenge dataset, focused on segmenting thoracic organs such as the heart, aorta, esophagus, and trachea. Built and evaluated PyTorch segmentation pipelines with multiple model options, loss functions, validation metrics, post-processing, and visualization tools for inspecting 2D slices and reconstructed 3D predictions. The project also included experiment tracking, reproducible training settings, and submission-ready prediction packaging.
Point-cloud processing project for turning LiDAR scans into planar representations. The tool is aimed at transforming `.e57` and `.ply` point-cloud files into detected planes, combining 3D geometry, scan processing, and documentation tooling. A compact project, but a nice example of working outside standard tabular or image data — closer to spatial data pipelines and real-world 3D reconstruction workflows.
Research project for the OpenDriveLab Autonomous Grand Challenge, extending SparseOcc with a custom flow-estimation module for predicting both 3D scene occupancy and object motion from six camera views. Our team reproduced the original SparseOcc model, integrated flow prediction directly into the sparse voxel decoder, and introduced a new multi-level MSE loss for supervising motion vectors. We evaluated the model on the nuScenes-mini dataset using RayIoU and Average Velocity Error, showing that flow prediction could improve occupancy performance under certain training settings. The accompanying technical report details the architecture, experiments, and lessons learned.
Group project on geometric deep learning for molecular generation, centered on Equivariant Diffusion Models for generating 3D molecular structures. The work combined a literature review, reproduction of the original EDM results from Hoogeboom et al., and a partial JAX reimplementation of the algorithm. The codebase includes QM9 data handling, EGNN components, diffusion-model logic, generated molecular samples, and a blogpost-style technical report explaining diffusion, graphs, equivariance, and why symmetry matters for molecules.
Research codebase for measuring and reducing gender bias in language models across eight languages: Dutch, English, French, Spanish, German, Korean, Japanese, and Chinese. The pipeline combines gender-neutral translation, LLM-based sentence completion, and bias evaluation with GAS, GLD, and ADD metrics. It supports both Hugging Face models and OpenAI models, with batch scripts for running multilingual experiments language-by-language to avoid GPU memory issues.
This site. Hand-coded HTML, CSS, and vanilla JavaScript — no frameworks, no build step. Served by nginx in a Docker container on my own Ubuntu homelab, tunneled to the public internet through Cloudflare, deployed automatically from GitHub on every push to main.
The full server environment everything else runs on: Docker, Portainer for container management, nginx reverse proxy, Cloudflare tunnel for safe public access, and enough networking decisions to fill a small whiteboard. A living project — the value is in the long-term maintenance, not the initial setup.
Vanilla survival Minecraft server running 24/7 in a Docker container, reachable through a custom subdomain. Includes DNS configuration, port forwarding, custom server properties, an automated backup routine, and a custom MOTD. Built for a small group of friends and maintained ever since.
Slowly going through low-level C++ exercises to keep my systems-programming instincts sharp. Small graphics demos, data-structure implementations from scratch, and the occasional shader. Counterweight to working in Python all day.