MSc Data Science student at ETH Zurich. BSc Artificial Intelligence from JKU Linz (with distinction), thesis on stabilizing autoregressive neural PDE surrogates. Previously Data Analyst at Tractive and Data Science Intern at Solvistas.

Interested in scientific machine learning, ML engineering, and building data products that people actually use.

• The Well (PolymathicAI, NeurIPS 2024 physics simulation benchmark)
  • #64 (merged): FNO and TFNO block outputs were computed and then discarded. Fix changed benchmark VRMSE from ~0.5 to ~0.3.
  • #63: LR scheduler state was not saved in checkpoints. Closed during a repo restructuring, maintainers asked for a resubmission after the release.
  • #78: long-time validation metrics were overwritten each batch. Confirmed by an independent impact study and fix pending.
  • #81 (merged): Windows path resolution for normalization files.

• BachelorThesis: benchmark of stabilization techniques vs. one-step training across 4 PDE datasets and 3 autoregressive architectures (FNO, U-Net, CNO). Built on The Well, with PDEBench converters. Thesis PDF in the repo.
• monster-angebote.at: scraper, website, and e-mail service for Monster Energy offers in Austrian supermarkets. About 100 Google Search clicks per week and about 100 e-mail subscribers. Flask, PostgreSQL, GitHub Actions, Render. Repo.
• PracticalWorkAI: conditioning (FiLM) and pushforward training for neural surrogates on turbulent radiative layers for rollout stability of autoregressive surrogates.

• 2nd of 143, molecule generation
• 4th of 134, molecule synthesis
• 2nd of 143, recommender systems
• 3rd of 273, image classification (custom CNN and training pipeline from scratch)

Python, PyTorch, SQL (PostgreSQL, Redshift), Hydra, Weights & Biases, Tableau, Docker, GitHub Actions, Flask.

Chess: FIDE 2000.