M.Sc. Computer Science (TU Berlin) building computer-vision and deep-learning systems, with a focus on neural 3D reconstruction and remote sensing.
Core stack: Python · PyTorch · Deep Learning · Computer Vision · LLM APIs
NeuroBathy: Physics-informed neural 3D reconstruction for shallow-water bathymetry from aerial imagery. Combines a modified NeRFrac and VGGT with a Snell's-law refraction correction; reaches 0.34 m RMSE against surveyed ground truth when a surveyed water surface is available.
Bathymetry-Fusformer: Reconstructs 0.25 m seabed imagery from 10 m Sentinel-2 data, guided by high-resolution bathymetry. Adaptation of the Fusformer transformer with an LMDB patch pipeline, masked training and GeoTIFF output.
MLCZ-Pipeline: Semantic segmentation of Local Climate Zones in Berlin, Milan and Athens, fusing 10 Sentinel-2 bands with 234 PRISMA hyperspectral bands. A U-Net (PyTorch Lightning) reaches 73.5 % test accuracy over 17 LCZ classes; includes cross-city transfer experiments. Explore the interactive city maps →.
food-facts-agent: LLM agent that answers questions about food products, such as "Which of these mueslis has the least sugar per 100 g?", with live Open Food Facts data via Claude tool use. Every answer cites products by barcode and names missing data; an automated evaluation of 15 questions with traps (unknown barcodes, missing values, ambiguous names) checks tool use, citations and every number in the answer, and 14 of 15 passed.