Mato Gudelj

Munich, Germany

About

I am pursuing a Master's degree in Computer Science at the Technical University of Munich. I currently hold a working student position in machine learning at Cruise, where I work on 3D detection in autonomous driving.

My main areas of interest are computer vision, generally applicable machine learning methods, foundation models and novel applications of deep learning in computer-aided human skill acquisition, as well as other problems. I am also interested in general software engineering, with a focus on high-performance code and C++.

I received my Bachelor's degree in Computing from the University of Zagreb in 2022. My thesis was in the area of computer vision. The main focus of the thesis was multi-domain semantic segmentation, where I developed a novel algorithm for deducing the sub/super/equal relationships between classes.

In 2024 I was selected as one of the winners of the OpenAI Preparedness Challenge, for providing novel ideas for catastrophic misuse of frontier models and societal harm prevention. This included a prize worth $25,000 in API credits. Find out more on the OpenAI blog post here .

Projects

(collapsible)

Vibesim.com

Vibesim landing page Vibesim login
Vibesim song view

An ML powered song recommendation website based on finding songs with similar vibes. It uses an audio embedding model with the database spanning over a million songs. The website is built with React, with the model side in PyTorch and the data and embeddings handled by PostgreSQL (pgvector). (click!)

Raytracer

Raytraced cornell box Raytraced power 8 Mandelbulb

A small modular raytracer written in C++. Along with the usual suspects, it includes SSE optimized sphere intersection calculation, raymarched implementation of the mandelbulb fractal and a portable JSON scene representation format. (click!)

Langly

Language website UI

A project aimed at making language learning more efficient. It is based on the N+1 language acquisition principle. The website aggregates target language (TL) material and orders it based on a quality score. Among other things, the score is based on the proportion of sentences that are N+1 and the frequency (usefulness) of the learned word(s).

SpriteGAN

A generated spritesheet containing 3x9 sprites

A DiffAugment StyleGAN2 trained on 32x32 sprites. The model was trained in early 2022. While the outputs are far from perfect and do exhibit partial mode collapse, to my knowledge this is the SOTA in equipment sprite generation at a fixed size. (click!)

Contact

Email will be loaded with javascript to prevent scraping.