ML engineer & researcher with a focus on Explainable AI (XAI) and medical imaging. I build systems that are both capable and accountable — interpretable classifiers, privacy-preserving local deployments, and agents that surface their own reasoning.
Explainability & Interpretability Concept Relevance Propagation (CRP), Layer-wise Relevance Propagation (LRP), prototype-based networks (ProtoPNet). My master's thesis explored concept-level attribution blended into a visual Web-UI, and pruning of image classifiers (which I inspected on the domain of medical diagnosis before)
Medical & Clinical AI Skin lesion classification (HAM10000), mitosis detection, and body-inclusive 3D scan processing for adaptive garment pattern generation using PointNet. Previously, I explored concepts and researched topics around histopathology and in the scopes of university courses in affiliation with Charité Berlin.
LLM Applications & Agentic Systems Local-first LLM deployment with Ollama, multi-persona conversational AI, privacy-preserving architecture for sensitive contexts. Interested in bringing interpretability thinking into the LLM space.
Full-stack ML FastAPI · React · Docker · TypeScript · PyTorch · TensorFlow
| Project | What it is |
|---|---|
ArXiv Paper Digest  |
LangGraph agent that monitors ArXiv daily — semantic filtering, local LLM summaries, SQLite memory. Available on PyPI. |
| Body Pattern Abstraction | PointNet-based sewing pattern adaptation from 3D body scans — disability-inclusive, privacy-first. Active research project. |
| RAG + XAI: Concept Pruning on Mitoses | RP-based concept attribution of a VGG mitosis classifier, extended with a RAG assistant that lets you query the XAI literature interactively |
| Political Chatbot Panel | Multi-persona AI debate panel, fully local (Ollama + Docker), built for a gallery exhibition |
| ProtoPNet ICW1 | Independent Coursework 1 (M.Sc. CS): Prototype-based skin lesion classification — 90.11% accuracy with interpretability preserved post-pruning |
- ArXiv Paper Digest — adding HuggingFace Space demo for zero-install access
- Body Pattern Abstraction — Phase 1: LandmarkNet training on synthetic scan data
- LLM fine-tuning (QLoRA) for domain-specific classification
- Mechanistic interpretability of transformer models
HTW Berlin · M.Sc. Computer Science