I am an AI/ML engineer and M.Sc. student in Artificial Intelligence & Intelligent Systems at the University of Bremen. I like projects where machine learning is not just trained, but measured, questioned, reproduced and integrated into a real system.
My strongest interests are:
- Medical AI — chest X-ray classification, robustness, explainability, fairness and calibration
- Robotics — ROS 2, navigation, planning, perception and autonomous systems
- Multimodal & biosignal ML — EEG, temporal signals and representation learning
- Applied AI systems — semantic search, retrieval, automation and research tooling
- Reproducible research — controlled experiments, provenance, raw predictions and honest evaluation
| Project | What I built / investigated | Highlight | Stack |
|---|---|---|---|
| MedVision-AI | Multi-label chest X-ray classification on NIH ChestX-ray14 with CNN backbones, asymmetric loss, XAI and planned SSL/fairness experiments | DenseNet-121: 0.7978 mean AUROC vs ResNet-50: 0.7067 across 14 pathologies in committed historical runs | PyTorch · timm · W&B · Grad-CAM |
| Gravloc | Spacecraft-hardware discovery and comparison prototype with datasheet extraction, semantic retrieval and grounded RFQ generation | Engineering-focused AI search with provenance and verification guardrails | Python · FAISS · Sentence Transformers · Streamlit |
| Intelligent Interactive Systems Robot Project | Mapping, localization, global planning and motion control for a mobile robot | End-to-end autonomous navigation stack | ROS 2 · A* · LiDAR · RViz · Python |
| EEG Motor Imagery Classification | EEG preprocessing and machine-learning classification experiments | Biosignal ML pipeline and comparative modelling | Python · EEG · Signal Processing · ML |
| Pneumothorax Shortcut Learning | Medical-imaging robustness experiments around shortcut learning in pneumothorax detection | Focus on whether models learn pathology or spurious cues | PyTorch · Medical Imaging · XAI · Robustness |
Research note: repository metrics are described as experimental results unless a linked paper explicitly states that the work has been peer reviewed or published.
MedVision-AI is being rebuilt as a publication-grade reproducible research project rather than just a collection of training scripts.
Backbones → ResNet-50 · DenseNet-121 · transformer baseline
Losses → BCE · Focal · Asymmetric Loss
Label budgets → 1% · 5% · 10% · 25% · 50% · 100%
Representation → ImageNet supervised vs medical self-supervised pretraining
Trustworthiness→ AUROC · AUPRC · calibration · subgroup analysis · XAI localization
Statistics → multiple seeds · bootstrap confidence intervals · saved raw predictions
The goal is not to chase one headline metric. It is to understand when a model improves, why it improves, and whether that improvement survives stronger evaluation.
I contribute fixes upstream instead of only working in my own repositories.
Project MONAI #8818 — merged fix using register_buffer for constant tensors in LocalNormalizedCrossCorrelationLoss, improving correct device and state handling in the PyTorch module.
You can also browse my pull requests and public repositories.
Languages
ML / Vision
Robotics / Engineering
research = {
"question": "clear and falsifiable",
"baseline": "reproducible",
"results": "saved with provenance",
"claims": "no stronger than the evidence",
"engineering": "clean enough to run again later",
}I enjoy turning rough experiments into systems that another person can actually understand, run and critique.
