Full Quest Log

This project was developed abroad in Denmark with three colleagues, inspired by a visit to Berlin. After visiting startups and state-funded labs that used data to solve urban questions, we were struck by the 'data footprint' of the city itself—specifically its street art. We wanted to see if we could treat graffiti not just as art, but as a complex system of spatial data. Using a DCGAN (Generative Adversarial Network), we built a model to synthesize urban typography like 'Wildstyle' and 'Bubble.' Beyond the technical challenge of balancing the Generator and Discriminator to avoid 'Mode Collapse,' the project was an exploration of how neural networks can interpret the creative 'chaos' of a city's visual language. It was our way of answering an urban question: can an algorithm learn the soul of a mural?

This project started as a solution to a security and performance bottleneck: legacy Windows 10 machines on campus were struggling with hardware limitations. I took two of these machines, wiped them, and installed Debian to build custom arcade stations. This led me down a rabbit hole of Linux tinkering—from fighting with outdated Wi-Fi chips that weren't compatible with modern signals (solving it through manual package injections and kernel tweaks) to bypassing restrictive campus firewalls to get the systems online. I also brought this home to my own Raspberry Pi, where I’ve been optimizing Batocera. For me, it’s about the satisfaction of understanding how the hardware actually talks to the software, and finding a workaround when the standard solutions don't work.

Developed with a colleague for a Big Data course, this project was born from our shared interest in books and a question: does the emotional tone of a review actually match its star rating? After finding existing Kaggle datasets were too small or full of missing values, we built a custom Python scraper to collect metadata and 16,000+ reviews for 4,000 titles. We processed the text using VADER sentiment analysis and found that numerical ratings often fail to capture the full nuance of user sentiment. Additionally, we modeled the data as a complex network using the Louvain algorithm, discovering that genres like 'Politics' and 'Children’s' act as essential bridges between different literary communities. The work involved intensive data cleaning and ANOVA testing, but the most interesting part was seeing how subjective human sentiments in comments are transformed into structured data and clear numerical results.