I am currently a master's student @ ETH Zurich and will be spending the spring and summer @ Stanford as a visiting student researcher.
I am interested in vision for robotics, generative modeling and scalability. I believe vision is the most important frontier with 70% of internet traffic being video and more than half of the human brain being devoted to vision processing.
Previously I completed my bachelors at ETH Zurich in electrical engineering, studied at NUS in Singapore and interned as a computer vision researcher at Polybee.
Fine-tuned Stable Diffusion on a custom corpus of "EasyRead data", pictograms that simplify texts for people with disabilities, and trained with specific user controllability in mind. Published at CHI 2026: https://dl.acm.org/doi/10.1145/3772363.3798916
I built a tool to semantically search for researchers and cluster knowledge for the Berkley AI Hackathon (2025). Earlier I developed a demo of how a marketplace in the future could look like using dense retrieval. Most recently I helped build the Stanford Registy embedding 10k out of 18k undergrad and grad students in over 100k vectors.
I came up with an emulated-double arithmetic solution (inspired by this 2006 paper) using paired single-precision (FP32) values to bypass the 64:1 FP32-to-FP64 performance penalty on consumer GPU architectures. The emulated doubles achieved 48 bits of effective precision and delivered a 3.8x speedup over native double-precision (tested on binomial option pricing). I wrote a blog post about this that got quite a bit of attention on hackernews.
Could we generate virtual worlds directly from brain signals? I was able to affirmatively answer this question by combining Nvidia Cosmos with DreamDiffusion to turn EEG readings into explorable environments.
Built at the dawn of the computer agent era in 24 hours and won Singapore's largest hackathon, Hack&Roll in 2025.
A novel architecture utilizing Monte Carlo Tree Search to generate interconnect topologies of diameter-3. This project, my bachelor's thesis, taught me RL and scalable engineering. The research on this project continues at the lab but this published paper uses my work.
This project taught me systematic ML experimentation: proper benchmarking, rigorous ablation testing, and drawing conclusive insights. Trained and evaluated a superresolution model based on this work on MRI images.
Backtracking, backtracking, backtracking! Got a bit obsessed with this puzzle and spent roughly 36 consecutive hours solving it.
I worked on the avionics stack of the 2023 ETH rocket team (ARIS). Launched our hybrid rocket at EUROC from a Portuguese army tank testing ground.
Trained a lightweight neural network, converted it to embedded C, and deployed it on an STM32 microcontroller for real-time song classification (~90% accuracy) using only a microphone and FFT-based features.
In my final year of high school, I became obsessed with understanding how China went from importing technology to dominating high-speed rail. Spent months diving into their strategy - tech transfer, absorption, leapfrogging. The thesis won provincial and national awards.