I'm a CS student heading to Soochow University for an AI Master's this September.
I work at the intersection of music information retrieval and embodied 3D safety understanding — both projects treat geometry (Tonnetz lattices; 3D Gaussian splats) as a first-class inductive bias for neural sequence / vision-language models.
| # | Project | Description | Tech |
|---|---|---|---|
| 1 | Chord-GPT | A Tonnetz-relative melody-to-chord harmoniser — my research repo; 65 unit tests, JOSS submission in prep |   |
| 2 | Echoesphere | My graduation project — an experimental interactive game where sound shapes the world |   |
| 3 | echoesphere_agent | AI decision-making agent for exhibition contexts |  |
| 4 | echoesphere-communication | Raspberry Pi sensor data collection & LED strip control |  |
| 5 | perry_hermes | A vibe-coded AI agent in Rust |  |
| 6 | amagicpear.github.io | My personal homepage |  |
| Domain | Stack |
|---|---|
| Languages |   |
| Frameworks |   |
| AI / ML |  |
| Music |  |
| Design |   |
| Infra |   |
- Exploring chord progression generation with geometry-aware tokenisation
- Exploring 3D safety understanding in 3D Gaussian Splatting scenes
- Building multimodal interaction systems (vision + language + hardware)
- Learning Rust for systems-level performance
- Making LLM-powered tools that simplify real-world workflows
"The best code is the code that doesn't exist — but if it must exist, make it minimal."
I'm not trying to write the most code. I'm trying to write the least code that does the most. Every refactor, every optimization, every architectural decision starts with one question: can this be simpler?
Built with ❤️ and minimal lines of code