1. dvit: that is 2.27 times faster than the standard vision transformer, adapts resolution at inference to switch between high and low resolution, trained on the cifar-10 dataset; no changes to be made in training and architecture.

2. rictr: knowledge distillation library for pytorch based on the papers: distilling knowledge in a neural network, paying more attention to attention: improving the performance of cnn via attention transfer and hidden state distillation; trying to reach on par with state of the art.

3. paper replications: from scratch in pytorch/jax, implemented deepseekv3, gemma, gpt, llama3, autoencoders, knowledge distillation, vision transformers and more.

4. raytracer: that i built to learn the math behind it; in just 99 lines of python, a cpu first ray tracer without acceleration that can render 3d models using triangle intersection and shading, resource used to build it is the ray tracing in one weekend book.

5. tinycpu: through my verilog circuit implementations, built the multistage 32 bit risc-v processor compiled (iverilog) and waveformed (gtkwave); has sub units for every module and an integrated unit for the processor with testbench.