At Blaize, I deploy CNNs including YOLO, ResNet, and PoseNet on the Graph Streaming Processor, and train segmentation models such as U-Net. I also developed on-device image downscaling and post-processing to preserve YOLO functionality while improving end-to-end efficiency.
I’ve ported models including Mask R-CNN to the GSP and developed parallel SIMD implementations for depth estimation and object tracking. I also created proof-of-concept implementations of small language models, including Phi-1.5, LLaMA 2, and Mixtral.
Previously, I optimized ThirdEye’s TensorFlow-based PoseNet pipeline and ported its SlowFast action detection pipeline to NVIDIA Jetson AGX Xavier. My work has also included C++/CUDA computer-vision and radar simulation, formal verification, and parallel simulation research.
