Omar Mohammed

Developing an FPGA-based neural network accelerator on AMD Versal hardware using Vitis AI and HLS, running model inference at far lower power than a CPU or GPU. Mapping convolutional and Mamba sequence models in C++/HLS and Verilog, including Mamba's sequence-dependent selective scan. Targeting low-power edge applications such as building energy monitoring, as part of an honors thesis.

Honors thesis in progress. Writeup and measured results to follow.

Custom 2-layer PCB for a battery-powered flood monitoring system. Integrates a FLIR Lepton thermal camera, BNO055 IMU, mDot LoRaWAN radio, and Witty Pi 4 power management for remote urban deployment.

First PCB I designed. Taught myself KiCad and went from schematic to fabrication-ready Gerbers in a few weeks. Mentored by Dr. Elizabeth Carter and Manu Shergill.

Estimates left/right sound direction from the Kinect 360's four-microphone array and fuses it with absolute heading from a servo-mounted IMU to point at the source in real time. Controller firmware in C, signal processing and a live Qt visualizer in Python, with telemetry over UDP.

Tracking is audio-only by design. There is no camera in the signal path, all processing runs on the host, and no audio is stored or transmitted.