AI Engineer
Bern, Switzerland
I design and build deep learning systems across research and industry. Currently pursuing my Master's in Artificial Intelligence in Medicine at the University of Bern.
My most significant work to date is PRISM, a generative AI framework I developed and co-authored for class-consistent image generation directly from structured tabular data, bypassing the language bottleneck in CLIP-based diffusion models. The work is under review at NeurIPS 2026 and achieves up to +116% improvement in class consistency over language-based baselines, with emergent geometric properties in the conditioning space arising without explicit supervision.
Alongside that, I've built vision-language architectures that outperform published baselines on zero-shot histopathology classification across seven public datasets, worked on brain lesion segmentation for longitudinal MRI in a clinical software product, and developed 3D path planning systems for autonomous underwater robotics. I enjoy tackling hard problems with intelligent systems and bringing research ideas all the way to working implementations.
AI Medical AG - Master's Thesis
Developing and benchmarking state-of-the-art deep learning architectures for brain lesion segmentation in longitudinal MRI scans and integrating the selected architecture into the company's Jazz software to enable automated lesion detection and progression tracking across patient follow-up studies.
Institute of Tissue Medicine and Pathology
Extended an existing vision-language framework by fine-tuning vision and language encoders on histopathology images and textual annotations, aligning their representations via contrastive loss for unsupervised downstream classification.
ARIS Space
Developed a ROS 2 node in Python for 3D trajectory generation and geometric path planning for the Nautilus autonomous underwater glider, computing turn radii, pitch angles, and waypoints.
Medical Park Goztepe
Diagnosed and repaired electrical circuits in medical devices to improve functionality and reliability. Monitored equipment performance and identified potential failures to ensure seamless operation.
Proposed a diffusion-based framework for generating fine-grained, class-consistent images directly from structured tabular data, bypassing the language bottleneck inherent in CLIP-based conditioning. Designed a two-stage training pipeline and inference strategy. Outperformed language-based baselines across three benchmarks: 0.96 on CUB-200 (vs. 0.83 FT-CLIP), 0.93 on PlantVillage (+116% vs. FT-CLIP), and matched Specialized CLIP on HAM10000 (0.73) using only structured metadata with no domain-specific pretraining. Discovered emergent geometric properties in the conditioning space without explicit supervision.
Developed six progressively improved models to generate captions for images. Used pretrained DINOv2 encoders and evolved decoders from RNNs to Transformers with cross-attention over spatial tokens. Evaluated with BLEU scores and visualized attention maps.
Designed and implemented a real-time machine learning pipeline for exercise quality assessment using WiiFit sensor data, including feature extraction, model training, and performance evaluation in a team-based setting.
University of Bern
Bern, Switzerland · Sep 2024 – Sep 2026
Deep Learning, Machine Learning, Computer Vision, From NLP to LLMs, Modeling & Scaling of Generative AI Systems, Reinforcement Learning, Trustworthy AI, HPC & Cloud Computing, C++ Programming
Bahcesehir University
Istanbul, Turkey · Sep 2020 – July 2024
Principles of AI, Linear Algebra, Differential Equations, Signals and Systems, Programming in Python & C, Biostatistics, Medical Imaging & MRI, Modeling and Simulation
I'm always open to discussing new opportunities, research collaborations, or interesting projects. Feel free to reach out!