I'm a CS Master's student at the University of Innsbruck, currently writing my thesis on LLM calibration. So I spend most days trying to make language models admit their mistakes instead of confidently hallucinating nonsense. I'm doing this research at JAIST in Japan, where I also co-authored a paper on step-wise confidence estimation.

I like building things that live in the terminal and implementing ML concepts from scratch because you don't really understand something until you've debugged it at 2am.

• LLM Calibration Research — extracting log-probabilities from Qwen 3 on GSM8K, AI2 ARC, and MMLU. Running on a single A100 that I treat with more care than my laptop.
• Published at JAIST — co-authored "Step-wise Decomposition Improves Calibration for Answering Multi-Hop Questions" (2025). Turns out telling an LLM to think step by step also makes its confidence scores better.
• IMC Prosperity 4 — placed top 25 globally in an algorithmic trading competition. Built ML-driven market-making strategies. My proudest moment was probably the late-night model rewrite that paid off.
• ML from Scratch blog — I implement ML concepts using only NumPy: gradient descent → linear regression → neural nets → Adam → LoRA. No frameworks allowed. Read here.