🔭 Interests: Machine learning, compilers, parallel computing, and embedded systems.

🌱 Recent Projects:

GPU Kernels: Metal Kernels for M1 GPU Mac

• Use Objective-C for host CPU to execute Compute Pipelines, and MSL (Metal Shading Langauge) for kernels
• Ex: 1D Prefix Sum, 1D Dot Product, 2D Matrix Multiply

Compilers: ChocoPy to RISC-V Code Generation + Optimization

• Wrote an end-to-end compiler in Java: lexing, parser, type checking, semantic analysis, RISC-V assembly code generation
• Implemented optimizations: constant folding, loop unrolling, in-lining, common subexpression elimination, and copy propogation

ML: PopDescent

• Build global optimization algorithm using Python’s TensorFlow+PyTorch to reach lower training loss on CNN/LSTM/LLM benchmarks
• Achieved 12% lower test loss in 15% fewer gradient steps against Sklearn, Keras Tuner, and popular learning rate schedules
• First author in “computational algorithm development” paper submission to ICML 2024; [https://arxiv.org/abs/2310.14671]

Firmware: Audio DSP at Johnson&Johnson MedTech

• Configured STM32 mcu to manually perform DSP and translate raw microphone bitstreams into encoded “AES” protocol
• Enabled circular buffering and manage interrupt race conditions for DMA memory transfer using callbacks and FreeRTOS
• Used Arm SIMD vectorization to optimize CPU memory by 15%, and decrease runtime by 8% to allow extra DSP functionality