Engineering Physics Student with Experience in Hardware, ML, FPGAs & PCBs
I am building a single-head scaled dot-product attention engine, the math behind every transformer, from a Python golden model down to synthesizable SystemVerilog, rebuilding the same operation at six levels of abstraction and diffing each against the last. The CUDA kernels hit float32 machine epsilon (1.19e-7); the RTL core is under way. A running dev-blog covers every milestone and bug.
Hardware Implementation of a neural network on a ZuBoard FPGA for inference. I managed to achieve 300% faster inference on my FPGA compared to my laptop for basic NIST dataset. Click below for more information.
Designed power boards, actuator boards, sensor boards, H-bridge boards & Wi-Fi communication for the winning robot in the UBC Engineering Physics Robot Summer competition.
I made an open-source IEEE 1500 Compliant Design For Test Block for Wearout Sensor IP developed by Prof. Andre Ivanov's SoC Lab. I got to work with Verilog & MultiSim to develop a compliant wrapper.
Using YOLO inference model and K-means, I was able to create a football tracking tool that takes in a clip and outputs an annotated clip that extracts basic statistics. This project showcases basic Computer Vision tools and will continue to be expanded.
I did research on the use of a MXene 2D film as a transistor using novel methodologies under Prof. Husam N. Alshareef. Publication is available on Advanced Materials Journal (29.4 Impact Factor) under the name: Wafer-Scale Transfer of MXene Films with Enhanced Device Performance via 2D Liquid Intercalation.