Cross-Domain Ontology-driven Agentic RAG over Gigantic Collection of German Historic Text running
Master Thesis · Expected Dec 2026

Building an agentic retrieval system over a large collection of German historical texts, combining a cross-domain ontology with graph- and hierarchy-aware retrieval to answer questions spanning literature, mythology, and the history of science.

Agentic Radiology Report Generation running
Master Thesis · Expected Nov 2026

Developing an agent-based pipeline that combines classification and segmentation models to automatically generate and evaluate radiology reports from CT data.

Conditional Generation of Tire Patterns using Generative AI running
Master Thesis · Expected Nov 2026

Fine-tunes a latent diffusion model with DreamBooth on tire tread imagery and guides generation with tire-performance evaluation functions, to automatically propose tread patterns meeting target performance characteristics.

Solving the Puppet Cube V1 with A* Search and Neural Networks running
Bachelor Thesis · Expected Oct 2026

Combining classical A* search with reinforcement learning to solve the Puppet Cube, a generalization of Rubik's-cube-style puzzles.

Automated Calibration of the Scan Trajectory in Dedicated Breast CT with Circle-Spiral Trajectory running
Bachelor Thesis · Expected Sep 2026

Automating the geometric calibration of a circle-spiral scan trajectory for a dedicated breast-CT system.

Self-Supervised 3D Masked Autoencoders for Volumetric CT Representation Learning running
Master Thesis

Extending masked-autoencoder pretraining from 2D to volumetric CT, studying cross-attention decoding and learnable register tokens for representation quality and efficiency.

Sampling-Aware Regularisation for Algebraic Reconstruction with Detector Gaps running
Master Thesis

Developing regularization strategies for algebraic CT reconstruction that account for gaps in detector sampling.

Noise Student-Teacher Training for Automated Labeling of Herculaneum Ink Detection Data running
Master Thesis

Using noisy student-teacher training to automatically generate ink-detection labels for the Herculaneum papyri scans, as part of the UnLost project.

Differentiable Compressed-Sensing ART
Project

Combined a differentiable ASD-POCS solver with a learned CT operator for compressed-sensing-style algebraic reconstruction.

Design and Evaluation of an LLM-Based Agentic Workflow for the (Semi-)Automation of Public Tendering Processes in Germany
Master Thesis

Designed and evaluated an LLM-based agentic workflow that automates large parts of German public-tender processing, from search and document analysis to cost calculation and bid submission.

Design and Evaluation of an LLM-Based Browser-Automation Workflow for German Public Tendering
Project

Builds a browser-automation layer for fragmented German e-procurement portals, where a language model discovers each portal's workflow once and a model-free executor replays it deterministically, reserving the binding bid submission for a human trigger; reaches 77-89% goal-state correctness on preparation tasks validated live on two real portals.

Deep Learning-Based Analysis of Thermographic Mold Images for Quality Prediction in Aluminum Die Casting
Bachelor Thesis

Trained CNN-based models on thermographic mold images to predict casting quality in aluminum die casting, using explainable-AI methods to highlight the most informative regions, in cooperation with an industry partner.

Deep Learning-Based Classification and Explainability of Cytomegalovirus Encephalitis in Longitudinal MRI Data
Project

Developed deep-learning models to classify cytomegalovirus encephalitis from longitudinal MRI scans, combined with explainability methods to highlight the anatomical regions driving each decision.

A Comparative Analysis of Deep Learning Models for Text-to-CAD Generation
Seminar

Compared two deep-learning approaches for generating 3D CAD models directly from text descriptions, showing a substantial reduction in invalid geometries with the newer transformer-based method.

Contrastive Learning of 3D Objects via Patch-Level Point Cloud Encoding for Similarity Matching
Master Thesis

Developed a patch-level contrastive learning framework for matching similar 3D objects from point-cloud data.

FVR-ADNeRF: Attention-Driven NeRFs for Few-View Reconstruction to enable CT Trajectory Optimization
Master Thesis

Explored attention-driven neural radiance fields for reconstructing CT volumes from only a few projection views, as a building block for trajectory optimization.

Generalization of Self-Supervised Vision Models in Image Retrieval
Project

Studied how well self-supervised vision models generalize across very different image domains for retrieval tasks.

Optimization of Image Reconstruction in Dedicated Breast CT with Circle-Spiral Trajectory
Master Thesis

Improved image reconstruction quality for a dedicated breast-CT system using a circle-spiral scan trajectory.

Design and Dataset Generation of Scanning Objects for CT Trajectory Optimization
Master Thesis

Built an automated, parametric pipeline for designing 3D test objects and generating a benchmark dataset for evaluating CT trajectory-optimization algorithms.

Image2TikZ: Neural Network-Based TikZ Code Generation
Master Thesis

Built a two-stage neural pipeline that converts diagram images into TikZ code, progressing from simple screenshots to hand-drawn sketches.

Diffusion Model-Based 3D CT Reconstruction for Arbitrary Trajectories
Master Thesis

Used diffusion models to denoise low-quality CT reconstructions from limited-view data acquired on arbitrary scanning trajectories.

Defect Detection Probability as a Metric for CT Image Quality Assessment
Project

Used the detection probability of a learned defect detector as a task-based image-quality metric for CT.

Automated ONNX2TikZ: Generating LaTeX-TikZ Diagrams of Neural Networks
Project

Automated the conversion of ONNX neural-network models into TikZ diagrams for use in papers and presentations.

Feature Extraction and Dimensionality Reduction Techniques for Assessing Similarity in Large-Scale 3D CAD Datasets
Master Thesis

Compared feature-extraction and dimensionality-reduction techniques for assessing similarity across large collections of 3D CAD models.

A Comparative Analysis of Loss Functions in Deep Learning-Based Inverse Problems
Project

Investigated how different loss functions affect the quality of deep-learning-based denoising and artifact reduction in low-dose CT reconstruction.

Developing and Evaluating Image Similarity Metrics for Enhanced Classification Performance in 2D Datasets
Bachelor Thesis

Developed new image-similarity metrics, combined with active learning, to improve classification performance on 2D image datasets.

Robot Movement Planning for Obstacle Avoidance using Reinforcement Learning
Master Thesis

Combined the artificial-potential-field method with reinforcement learning for collision-free robot-arm motion planning. This work led to a publication in Scientific Reports.

Detectability Index Reimplementation for CT Images Using PyTorch
Project

Reimplemented the detectability-index method for evaluating individual CT projections in PyTorch, with a focus on performance and maintainability.

Scamming Scammers Using Large Language Models
Master Thesis

Studied how large language models can be used to automatically engage and waste the time of email scammers, including how to estimate the resulting economic cost to the scammers.

Deep Learning Computed Tomography based on the Defrise and Clack Algorithm for Specific CBCT Orbits
Master Thesis

Investigated learned, trajectory-specific filters for FBP-style reconstruction on non-circular CBCT orbits. This work laid the groundwork for a series of follow-up publications on shift-variant FBP reconstruction.

Sinogram Analysis Using Attention U-Net: A Methodological Approach to Defect Detection and Localization in Parallel Beam Computed Tomography
Master Thesis

Developed a three-stage attention U-Net pipeline that detects and localizes defects directly in the sinogram domain, without reconstructing an image first. This work led to a publication at the CT Meeting 2024.

Learning Reconstruction Filters for CBCT Geometry
Project

Explored iterative and learned-filter approaches for CBCT reconstruction, including a data-driven FDK filter that generalizes the classical Ram-Lak filter.

Application of Projection-Based Metrics to the Optimization of Arbitrary CT Scanning Trajectory
Master Thesis

Compared a range of projection-based metrics for optimizing CT scanning trajectories and used a learned sequence model to search for improved trajectories. This work led to a publication at the CT Meeting 2024.

Tackling Traveling Salesman Problem with Graph Neural Network and Reinforcement Learning
Project

Solved the Traveling Salesman Problem using graph neural networks trained with reinforcement learning, comparing actor-critic and double-Q-learning training.

Synthetic Projection Generation with Angle Conditioning
Project

Synthesized CT projections at previously unseen viewing angles from a handful of known-angle input projections, aiming to reduce radiation exposure.

How to Learn and Use the Detectability Index Efficiently for CT Trajectory Optimization
Bachelor Thesis

Studied how to learn and use the detectability index efficiently to guide CT trajectory optimization.