Designed and implemented a GPU-accelerated MNIST training engine fully in WebGPU.
Built complete forward and backward passes using 6 custom WGSL compute shaders:
Dense layer forward kernels
ReLU activation
Softmax + cross-entropy backward
Dense layer gradient kernels
SGD optimizer
Implemented GPU memory layouts, buffer management, shader orchestration, and dispatch scaling.
Achieved ~40% training accuracy after 20 epochs on batch 64 in browser.
Demonstrates ability to build GPU compute pipelines and AI runtimes from low-level operations.
