My research explores how AI models understand integer sequences and their underlying logic. As part of my bachelor’s thesis, I collaborated with two Ph.D. students from ETH Zurich to develop an enhanced Transformer model that improves AI’s capacity for mathematical and logical reasoning, an area where existing models struggle. Our adapted architecture was designed to learn the abstractions within integer sequences, bridging the gap between language understanding and mathematical reasoning. This work has been published at NeurIPS, one of the most esteemed conferences in AI. To support our research, we created a comprehensive benchmark and dataset, along with a full suite of tools for training and testing models on integer sequences.
As a Research Assistant at ETH Zurich, I advanced research on super-resolution tasks using Implicit Neural Representations (INRs), building upon my Master's thesis work. Unlike traditional digital methods that store data in discrete grids with fixed resolutions, INRs model data as continuous functions, enabling sampling at arbitrary resolutions. This approach is particularly well-suited for super-resolution tasks. To further explore their potential, I trained a Vision Transformer model on diverse data modalities, demonstrating the versatility of INRs. Our findings are on track to be published at ICCV, one of the leading AI conferences.
CodebaseTogether with two fellow ETH graduates, we developed a comprehensive knowledge retrieval system from scratch for the Swiss bank Julius Baer, leveraging large-language models fine-tuned on internal documents. This system enhances employee productivity by enabling fast and efficient access to company information and policies while ensuring compliance. Built with Python and PyTorch, it seamlessly integrates into daily workflows. To further optimize document retrieval, we constructed an embedding space, enabling the model to provide accurate source citations when delivering information.
Julius Bär
In addition to developing AI projects in Python,
I have a strong interest in frontend development
with HTML, CSS, and JavaScript. This passion has
led to several web design projects, some dating
back to my high school years. Notably, I built
my own website entirely from scratch using only
HTML, CSS, and JavaScript, without any
additional frameworks or libraries.
My
first complete web development project was in
2014 for Thurnheer Obst, a small family-owned
agricultural company specializing in fresh local
fruits. Although I no longer maintain the site,
it marked my introduction to web design. I later
contributed to the web design for airlinetools,
a company that developed a sophisticated
web-based Airbus A320 simulator built by
professional pilots.
Most recently, I designed and currently maintain
the website for Avanox AI, an AI startup where I
am also a co-founder. At Avanox AI, we simplify
AI-driven automation for businesses of all
sizes.
I have a deep passion for cryptography, blockchain technology, and cryptocurrencies. My interest in distributed computation was sparked during my master's studies through multiple courses from my supervising professor, whose work on improving the Bitcoin Lightning Network was both inspiring and influential. As part of this passion, I developed a cryptocurrency trading bot that directly interacts with the Solana and Ethereum blockchains, utilizing Node.js, Rust, and Solidity. Additionally, before Ethereum’s transition to Proof of Stake with ETH 2.0, I built and sold various Ethereum mining rigs.
Trading Bot