Résumé 

Student Researcher

Visual Geometry Group at University of Oxford

Exploring the limits of 3D reconstruction.

Research Assistant

Database Systems, Data Mining and AI Group at the Ludwig Maximilian University of Munich

At the group I conducted research on multidimensional analysis of the topography and features of data clusters with the aim of determining suitibility of different clustering algorithms.

• Clustering Algorithm Analysis
• Data Topography

Applied Science Dual Student

AI Labs - Amazon Web Services (AWS)

The science of scaling large language models is an interesting field of research. For me it was shaped by combining aspects of graph theory with the transformer architecture to develop a new powerful pruning method.

• Inference Optimization
• Graph Theory
• Transformers

Applied Science Dual Student

SageMaker - Amazon Web Services (AWS)

I am working on the science aspects of improving and developing Amazon SageMaker automatic model tuning, a fully-managed platform enabling data scientists and developers to easily build, train, and deploy machine learning models at scale.

• Hyperparameter Optimization
• Auto ML
• Forecasting

Data Science Dual Student

ProServe - Amazon Web Services (AWS)

In my time at the Professional Services Emerging Technologies team I worked on the identification of topic trends in tech with relevance to the organization using Natural Language Processing. I worked with Huggingface Transformers as well as variety of different conventional NLP models such as Latent Dirichlet Allocation and Term Frequency - Inverse Document Frequency. I also got the chance to shadow a ProServe customer project, getting insights into the daily life of a consultant.

• Natural Language Processing
• Forecasting

Data Science Dual Student

Account Management - Amazon Web Services (AWS)

In my time at the Professional Services Emerging Technologies team I worked on the identification of topic trends in tech with relevance to the organization using Natural Language Processing. I worked with Huggingface Transformers as well as variety of different conventional NLP models such as Latent Dirichlet Allocation and Term Frequency - Inverse Document Frequency. I also got the chance to shadow a ProServe customer project, getting insights into the daily life of a consultant.

• Churn Modelling
• Data Driven Sales

First aider in the response unit

Bavarian Red Cross

Master of Science (Grade 1.0*)

Ludwig Maximilian University of Munich

Elite Data Science Master

• Statistical Reasoning and Inference
• Big Data Management and Analytics
• Algorithm Design
• Advanced Statistical Modeling
• Supervised Learning
• Connecting Language and Vision
• Introduction to Deep Learning
• Data Security and Ethics
• Human Computation and Analytics

Term Abroad (Grade 1.1*)

University College Cork

Computer Science

• Network Science: Theory and Applications
• Theory of Computation
• Artificial Intelligence I
• Computational Machine Learning
• Big Data Management and Analytics
• Strategy for Global Organisations
• Principles of Management and Organization

Bachelor of Science (Grade 1.2*)

Baden-Wuerttemberg Cooperative State University

Business Information Systems - Data Science

LLM-Rank: A Graph Theoretical Approach to Pruning Large Language Models

arXiv

This work introduces MLPRank, a pruning method that applies graph-theoretic centrality—specifically a modified weighted PageRank—to identify important nodes in multilayer perceptrons, enabling structured sparsity and reduced inference cost. An extended version for transformer decoders, LLMRank, is also proposed. Both methods deliver strong results: MLPRank retains 6.09% more accuracy than baseline pruning approaches, and LLMRank improves accuracy retention by 13.42% over leading transformer-based baselines.

Autoencoder-based General Purpose Representation Learning for Customer Embedding

arXiv

This work introduces DEEPCAE, a new regularization method for deep contractive autoencoders designed to embed complex tabular entities. Using a general-purpose embedding framework, the approach outperforms other autoencoder variants in both reconstruction and downstream prediction tasks. Across 13 datasets, DEEPCAE reduces reconstruction error by 34% compared to a stacked CAE.

Impact of HPO on AutoML Forecasting Ensembles

arXiv

This work evaluates how adding hyperparameter optimization to deep-learning-based forecasters (DeepAR and MQ-CNN) affects the performance of a diverse forecasting ensemble. The study shows that tuning can significantly improve accuracy—boosting avg-wQL by 9.9%—but also increases end-to-end latency by 65.8%. Using Bayesian Optimization and Hyperband, the final configuration outperforms Amazon Forecast, achieving 3.5% lower error with 16% lower total latency.