Piano × machine learning × hardware.
I turn messy real-world signals — MIDI files, camera frames, circuit boards — into models that make a useful call. Below are four projects taken end to end: from raw data to a trained model to something you can click, hear, or watch move.
Predicting how hard a piano piece is to play, from the MIDI alone, across the 11 grade levels used by the ABRSM, RCM, and Trinity syllabi. A gradient-boosted model on hand-designed features beats both a Transformer I also built and a published deep-learning benchmark, landing near the task's label-noise floor.
| Parameter | Value |
|---|---|
| Feature-model MAE | 0.959 |
| Transformer MAE | 1.131 |
| Label-noise floor | ~0.7 |
| Difficulty grades | 11 |
[ scikit-learn · Gradient boosting · PyTorch · PSyllabus ]
→ Explore the modelYOLO11 instance segmentation finds 20 classes of electronic components on a circuit board, maps pixels to printer coordinates, and physically picks and sorts them with a vacuum gripper on a modified 3D printer. Synopsys Science Fair 2026.
| Parameter | Value |
|---|---|
| Box mAP@50 | 84.5% |
| Mask mAP@50 | 84.0% |
| Classes | 20 |
| Images | 7,559 |
[ YOLO11 · Segmentation · Coordinate transforms · Pick & place ]
→ See the pipelineA still painterly pond that listens: piano notes scatter ripples across the water, loudness brings light, harmony tints the sky, and sustained playing warms the scene from night to dawn. An independent piece written in ChucK with the ChuGL graphics engine; the page runs a WebGPU port live in the browser.
| Signal | Response |
|---|---|
| Note onset | ripple |
| Loudness | light |
| Harmony | sky hue |
| Sustained playing | dawn |
[ ChucK · ChuGL · WebGPU · WebChucK ]
→ Enter the gardenThe C-major "first light" passage of Liszt's Vallée d'Obermann as a living alpine lake: notes ring the water, bass chords throw meteors, and light floods in only as the harmony resolves. The actual hand-written ChucK program runs in the browser — the real ChucK VM and ChuGL engine compiled to WebAssembly, playing itself.
| Signal | Response |
|---|---|
| Note onset | ripple |
| Harmony resolving | first light |
| Bass chord | meteor |
| Musical energy | koi & birds |
[ ChucK · ChuGL · WebAssembly · AudioWorklet · WebGPU ]
→ Stand in the first lightClassical pianist for 5+ years, with performances at Carnegie Hall in 2024 and 2025. Select any line to watch the recording.
| Domain | Tools |
|---|---|
| AI / ML | PyTorch, Transformers, YOLO11, Computer Vision, Instance Segmentation, Ordinal Regression |
| Music Tech | MIDI Analysis, Feature Engineering, Gradient Boosting, AdamW, PSyllabus Dataset |
| Software | Python, TypeScript, OpenCV, CLAHE, Mixed Precision (FP16) |
| Hardware | Arduino, Klipper Firmware, 3D Printing, Stepper Motors, BLTouch, Vacuum Pickup |
Strip the pitches from a piano performance — keep only timing, dynamics, and pedal — and can a model still tell Chopin from Bach? On 30-second MAESTRO excerpts across five composers, our gradient-boosted trees reached 0.61 macro-F1 held-out against 20% chance. The bigger lesson: the official MAESTRO split leaks works across train and test, so we rebuilt composition-level splits and kept the test set sealed.
→ Learn moreAn independent deep-dive into ChucK and ChuGL: two hand-written real-time pieces where sound and image come from one program. Resonant Garden — a pond that wakes with live piano — runs as a WebGPU port with WebChucK audio; Pèlerinage goes further, running the actual ChucK VM and ChuGL engine compiled to WebAssembly in your browser.
→ Enter the gardenI assumed a Transformer would win at predicting piano-piece difficulty. It didn't: a gradient-boosted model on hand-designed features — note density, hand independence, leaps — beat my 2M-parameter Transformer and a published benchmark, near the label-noise floor. An ordinal loss appeared to collapse along the way, but controlled experiments couldn't reproduce it.
→ View projectStrip a 3D printer to its gantry, bolt on a vacuum nozzle and a BLTouch probe, and let YOLO11 do the seeing. The hard part wasn't the network — it was the camera-pixel-to-millimeter coordinate transform, solved with a degree-3 polynomial mapping and sub-millimeter probing.
→ View projectFour open-source PCB datasets, 48 raw class labels, zero consistency. Weeks of merging, remapping, and validating annotations produced a unified 20-class dataset with 7,559 images — and 84.5% box mAP, 84.0% mask mAP at IoU 50.
→ View project