Saurav Paudel

At LogicTronix Technologies, I build real-time perception pipelines for ADAS on Xilinx Kria K26 hardware. I deployed a Prophesee event-camera person-threat detection pipeline using YOLOv8n INT8, reaching 37.8% mAP@0.5 at 35 ms end-to-end latency.

I've accelerated PointPillars preprocessing from NumPy to C++/OpenMP with Pybind11, reducing runtime from roughly 250 ms to 25 ms on Kria ARM Cortex-A9. I also improved NanoDet sign AP@0.5 from 14% to 35% by projecting Waymo 3D LiDAR sign labels into synchronized multi-camera views.

I work across model design, compression, quantization, and bare-metal deployment. My work includes redesigning YOLOv2/v3/v4 for Xilinx DPU compatibility, building VitisAI INT8 quantization and fast-finetuning pipelines, and developing lightweight PetaLinux inference post-processing.

Outside edge deployment, I build high-performance C++/CUDA attention kernels for Ampere and Hopper GPUs, using SRAM tiling, Online Softmax, Tensor Cores, asynchronous data movement, and Nsight Compute profiling.