CV

Research Assistant

Friedrich-Alexander-Universität Erlangen-Nürnberg

• Proposed an anisotropy-based metric for detecting and mitigating memorization in diffusion models (Stable Diffusion), ~5–8× faster than prior methods and requiring only 2 forward passes (ICLR 2026); also researching generalization in VLMs and VLA models.
• Developed DiNAS, a conditional graph diffusion model generating neural architectures in <0.2s across 6 benchmarks (TMLR 2024), and Dextr, a zero-shot NAS proxy using SVD and extrinsic curvature requiring only one unlabeled sample (TMLR 2025).
• Collaborating with Nokia and Astrum IT on remaining useful life (RUL) prediction of hardware components using pretrained time series foundation models.
• Supervised ~15 MSc theses and 2 seminars; co-wrote DFG research proposals.

Machine Learning Intern

BMW Group

• Predicted Quality of Service attributes (ping, throughput) using large-scale real-world signal data with up to 7 million samples through CNNs, SVMs, and XGBoost, achieving state-of-the-art accuracy on ping prediction.

Working Student

Fraunhofer Institute for Cognitive Systems IKS

• Built a reverse image search system using transfer learning embeddings and t-SNE to retrieve visually similar products with metadata (price, condition, material) from the web.

PhD

Friedrich-Alexander-Universität Erlangen-Nürnberg

Machine Learning (Chair of Multimedia Communications and Signal Processing)

M.Sc.

Ulm University

Cognitive Systems

B.Tech.

Dr. APJ Abdul Kalam Technical University

Computer Science & Engineering

Detecting and Mitigating Memorization in Diffusion Models through Anisotropy of the Log-Probability

International Conference on Learning Representations (ICLR) 2026

Dextr: Zero-Shot Neural Architecture Search with Singular Value Decomposition and Extrinsic Curvature

Transactions on Machine Learning Research (TMLR), 08/2025

Multi-conditioned Graph Diffusion for Neural Architecture Search

Transactions on Machine Learning Research (TMLR), 03/2024 (25 citations)

Differentiable Cost Model for Neural-Network Accelerator Regarding Memory Hierarchy

IEEE

Too-Hot-to-Handle: Insights into Temperature and Noise Hyperparameters for Differentiable Neural-Architecture-Searches

IEEE

PSumSim: A Simulator for Partial-Sum Quantization in Analog Matrix-Vector Multipliers

IEEE

Fake News Detection using Deep Learning Models: A Novel Approach

Transactions on Emerging Telecommunications Technologies