Research, side projects and more

Berkley Faculty Network - a visual graph to find all people relevant to your research

• • Built a tool for the Berkley AI Hackathon with Daniel Kiss to semantically search for researchers, letting you explore the ones most relevant to your interests through an accessible visual graph.
• • Starting with 14 Berkeley faculty leading in AI, we identified a total of 12,336 researchers within just two degrees of separation - based on publications' co-authorship - and mapped them all visually and semantically.
• • Take a look here.

Car Finder - a car marketplace using vector search

• Built this demo of how vector search could change marketplaces together with Daniel and Varun.
• • We used the text-embedding-3-small model from Open AI to generate the embeddings and cohere for reranking of search results. 4o-mini served as the LLM model for rewriting queries and to augment our user facing retrieved results.
• • Check it out here.

Hack&Roll 2025 - built a computer agent that helps non-tech-savvy people navigate their desktops

• Built an agent that provides step-by-step, visual guidance by spotlighting where users should move their cursor.
• • Built in 24 hours at Singapore's largest hackathon and won in the main category.
• • Here are the links to the devpost entry and our github repo.

Bachelor's Thesis - A Reinforcement Learning Approach to Generate Network Topologies

• Developed an architecture from scratch utilizing Monte Carlo Tree Search to generate interconnect topologies of diameter-3 for data centers at the Scalable Parallel Computing Lab of ETH Zurich.
• • The thesis is not published as of now, but part of my work served as foundation for this paper (see acknowledgments).
• • Profiled, optimized and parallelized the architecture for a 10x increase in self-play games.

Superresolution using LIIF

• Improved upon the existing model by Chen et al for Superresolution using Local Implicit Image Functions.
• • Performed quantitative cross validation tests over medical datasets leading to improved PSNR results.
• • Adjusted and trained existing model to allow for single image training leading to a strong color shift but reasonable generalization.
• • Used slurm to train new model on a dedicated GPU cluster.

March 2024 Jane Street Puzzle

• Solved the Jane Street Puzzle of March 2024 algorithmically and submitted my solution.
• • Implemented algorithm to backtrack through all possible combinations using C++.
• • Optimized algorithm to find the solution in 65'536 hook positions in under 20 minutes on my personal laptop.

Bohnanza Cardgame

• Implemented the card game Bohnanza for the very first time.
• • Set the requirement elicitation and the software design requirements.
• • Designed and Implemented frontend using the WxWidgets framework in C++.
• • Conceived a scheme to allow for interplayer trading of cards.
• • Streamlined project coordination using Git.

BERNOULLI student rocket

• Built a hybrid rocket with an autonomous guided recovery system in a student team and participated in the European Rocketry Challenge (Lisbon, Portugal) in the 9000m category.
• • Overcame challenges in working with undocumented code of the previous team to implement and test our sensor drivers.
• • Closed multiyear sponsoring agreements with industry partners.

Mini-Shazam

• Designed and implemented a model to recognize songs on a low-power microcontroller.
• • Captured song snippets on the microcontroller, trained them on an external computer and tested the model on the microcontroller (used the leds to encode each song).
• • The model on the microcontroller was able to classify songs correctly with an accuracy of about 85% on our dataset.

Nominated Thesis on the Chinese HSR system

• My high school thesis on China’s remarkable progress towards becoming the global leader in railway manufacturing won a number of awards.
• • Price in the National Competition by Schweizer Jugend Forscht.
• • Third Place at the 2. Future Skills Forum Rheinfelden.
• • State Level Award by the Canton Zurich.