may – august 2025 • los altos, ca

• built matmul kernels on blackwell that beat nvidia's cublas at 103% sota (blog, code)
• wrote 1d conv kernel that beat cudnn by 7x, speeding up inworld ai's tts model by 12%
• implemented hilbert curve scheduling for h100
• built specialized kernels (3d conv, bicubic resize) for midjourney inference

sep – dec 2024 • fremont, ca

• built computer vision prototype for 2-post lift safety in service centers
• trained segmentation model to detect lift arms and compute safe angle thresholds

jan – april 2024 • dearborn, mi

• built automated test suite using playwright and typescript

may – aug 2023 • waterloo, on

• built full-stack analytics dashboard for service data

cuda • github

• built cuda matmul kernel for a100 from scratch, progressively optimizing performance
• implemented memory coalescing, increasing bandwidth from 15gb/s to 110gb/s (6.5× speedup)
• added 1d block tiling to improve register reuse, achieving 8.5 tflops

python, numpy • github

• built neural network from scratch using only numpy
• implemented dense layers, activation, and loss functions
• built custom optimizers (sgd, adam) and mini-batch gradient descent