Jay Shah

A GPU profiler that instruments compute shaders and tracks execution metrics — memory usage, workgroup efficiency, and shader timing — without requiring any changes to the application being profiled. WebSight gives developers visibility into GPU behaviour that is otherwise opaque. It includes memory leak detection, shader complexity checks, and workgroup optimization analysis, surfacing actionable insights to improve throughput and stability.

Accelerated a pretrained chest X-ray pneumonia detection model on NVIDIA A100 GPUs. The goal was to reduce inference latency without sacrificing model accuracy. By applying FP16 mixed precision, layer-wise quantization, and operator fusion, and using NVIDIA Nsight and TensorRT to profile and eliminate bottlenecks, the project achieved a 50% reduction in end-to-end inference time with improved memory efficiency and throughput stability.

An analytics platform that runs SQL-style queries directly on the GPU against multi-million record datasets. GridDB implements parallel SELECT, WHERE, GROUP BY, ORDER BY, and histogram operations using CUDA and WebGPU kernels, with both execution paths profiled for occupancy, synchronization overhead, and latency. Results are rendered as scatter plots, histograms, time-series charts, and KPI dashboards — enabling fast visual insights for quality and process analytics.