Notes on research, engineering, and the creative process
AI/MLMusicResearch
May 2026
Why CORAL Fails on Piano Grading — and How CORN Fixed It
When I trained a Transformer to predict piano-piece difficulty, a popular ordinal regression method (CORAL) completely collapsed. After weeks of debugging, I discovered that CORAL's shared weight structure can't handle the subtle, overlapping features that separate adjacent difficulty grades. Switching to CORN — which gives each grade its own classifier — boosted exact-match accuracy by 9.4%. Here's the full story of how I found the problem and why CORN works.
Turning a $200 3D Printer into a PCB Component Sorter
E-waste recycling plants lose millions of reusable components because manual sorting is too slow. My solution: strip a 3D printer down to its gantry, bolt on a vacuum nozzle and a BLTouch probe, and let YOLO11 do the seeing. The hardest part wasn't the neural network — it was the coordinate transform from camera pixels to millimeter-accurate pickup positions. A degree-3 polynomial mapping and sub-millimeter probing got us there.
Four open-source PCB datasets, 48 raw class labels, zero consistency. Capacitors were labeled 'cap', 'C', 'capacitor', and 'SMD_cap' across different sets. I spent weeks merging, remapping, and validating annotations to create a unified 20-class dataset with 7,559 images. The result: 84.5% box mAP and 84.0% mask mAP at IoU 50. Data engineering isn't glamorous, but it made or broke this project.
This Summer: Real-Time Audiovisual Composition at Stanford CCRMA
I'll be at Stanford's Center for Computer Research in Music and Acoustics this June for a five-day workshop on RayTone — a node-based environment that lets you compose sound and visuals simultaneously. I'm especially excited about learning ChucK for audio synthesis and GLSL shader programming for graphics. The workshop culminates in an original multimedia composition, and I'm planning to explore how AI-generated music features could drive real-time visuals.
This Summer: Audio-Centric Game Design with ChucK & ChuGL
Right after RayTone, I'm coming back to CCRMA in August for a second workshop — this time building interactive audiovisual software with ChucK and ChuGL. The goal is to prototype a video game where sound isn't just decoration but a core mechanic. Taught by the creators of ChucK themselves, including Professor Ge Wang. I want to combine what I've learned about music difficulty grading with interactive game design — imagine a rhythm game that adapts its difficulty in real time.
This Summer: UC COSMOS at UCSD — Machine Learning from the Ground Up
I've been selected for UC COSMOS Cluster 7 at UC San Diego: 'Machine Learning: Can We Teach a Computer to Think?' It's a four-week residential program at the Jacobs School of Engineering where we'll build ML systems from scratch — not just call APIs, but understand the math and intuition behind how models learn. After spending months training Transformers and YOLO networks for my own projects, I'm excited to fill in the theoretical foundations and work alongside other students who are just as obsessed with this stuff.