PhD student in Machine Learning & NLP at the University of Tübingen, and Tech Lead of Scholar Inbox.
I am the Tech Lead of Scholar Inbox, a personalized paper-recommendation platform now used by 40k PhD-level researchers. I lead a team of 10 and built the recommender system, the active-learning onboarding that solves the cold-start problem, and the semantic search and retrieval stack.
I am a PhD student at the Autonomous Vision Group in Tübingen, supervised by Prof. Andreas Geiger. I work on natural language processing: long-context transformer models, generative recommendation systems, and the evaluation of research ideation agents.
My path began with Theoretical Physics at Imperial College London, followed by Data Science at Marburg, a Master's in Computer Science in Tübingen, and 8 months as a data scientist at Doxa Partners LLP in London.
The first comparable evaluation technique for research ideation agents that goes beyond LLM-as-a-judge, grounding evaluation in future scientific discoveries.
A long-context, hierarchical transformer-encoder with a custom CUDA kernel giving a 7.5× training speed-up.
Lead the technical development of scholar-inbox.com, coordinating a team of 10. Built the recommender system, active-learning onboarding, semantic search & retrieval, and the UI. 40k PhD-level users to date.
First work on Scholar Inbox with Python, PostgreSQL and Flask. Scaled the platform from one to a few hundred users while obsessing over customer experience.
Built “Chronicle”, a no-SQL time-series database with state-of-the-art write speeds, now in use at the CERN particle accelerator. Iterated on the Java API and refactored a complex inheritance structure.
Supervised by Prof. Andreas Geiger. Published a long-context hierarchical transformer with a CUDA kernel, a recommendation system and dataset, and grounded evaluation for research ideation agents.
Master's thesis on recommender systems for Scholar Inbox, reaching 96% NDCG. Supervised directly by Prof. Andreas Geiger.
Bachelor's thesis on predicting antimicrobial peptides using evolutionary algorithms. Learning: don't use evolutionary algorithms.
