comp eng @ waterloo · first year
I like making things run on hardware they weren't built for, or faster than designed. Specifically, I wrote Metal kernels that sped up a detector 1.8x and got RDR2 running on macOS. Looking for a winter 2027 co-op (jan-apr).
Looks at your shooting form and gives feedback. It can also rate shots in any game from 1v1 to 5v5 by analyzing separation. I optimized the GPU kernel and video pipeline to cut processing time by 88% on a test clip.
python / pose estimation / ml
Rewrote ballform's ball detector in MLX with custom Metal kernels. It's 1.8× faster than ONNX Runtime + Core ML on an M3 Pro with the same detections, and preprocessing went from 4.7 ms on the CPU to 0.42 ms on the GPU.
python / metal / mlx / gpuAn iPhone widget that shows my Claude session and weekly usage, so I always know how much I have left. A menu bar plugin on my Mac syncs the numbers through iCloud every 5 minutes.
javascript / scriptable / swiftbar
Got RDR2 running on my Apple Silicon Mac, which it has no official support for. I put together a CrossOver Wine and Apple D3DMetal setup, patched Wine to fix crashes in Rockstar's launcher, and wrote the scripts that start it.
c / python / wine / metal
A 5-DOF character lamp that uses a person's arm movement to control the simulated joints.
python / computer vision / simulation
A tool for filling out annoying PDFs, now used by 70+ people a month. I trained a D‑FINE object detector on 13K+ annotated forms to find fields in flattened PDFs, then built the pipeline that pairs it with language models to fill them in.
typescript / react / python / pytorch / postgresqlElectrical and programming for our 30+ person robotics team. One project was a targeting algorithm that let the robot line up a shot while it was moving. We won the 2026 CA District Glendale event.
Built a few mid-drive and hub-drive e-bike conversions. Made wiring harnesses, integrated the electronics, and spent a lot of time debugging battery problems.
Vision and language models on Apple Silicon (5.5K+ GitHub stars). I sped up and fixed its RF-DETR object detector, checking every change against the original PyTorch model.
Apple's array framework for machine learning (28K+ GitHub stars). I found that conv2d gave different results for the same image depending on batch size, because a faster but less precise algorithm kicks in from a batch of 3.
SoCal Rehab — ML Engineer
Automated clinical paperwork with Vertex AI and Apps Script. Cut processing for a 100+ patient batch from about 5 hours to 15 minutes.
AutoFillPDF — Founder / Engineer
Built and launched a document automation platform with 70+ monthly users. Did the product, the ML pipeline, and everything else.
Python, TypeScript, C++, Java, SQL · PyTorch, MLX, Core ML, ONNX Runtime, OpenCV, MediaPipe · Metal, FastAPI, Next.js, PostgreSQL, Vertex AI, Linux, Arduino, Raspberry Pi