Check out my Github profile for a complete list of my projects.

▸ Multi-region-attention
Multi-region-attention Multi-region-attention

github

MRA adds multi-scale regional self-attention to Vision Transformers, letting each pixel attend to coarse-to-fine neighbourhoods. It slashes quadratic cost while matching SOTA segmentation accuracy on Cityscapes and BraTS.

▸ Color Constancy Project
Color Constancy Project Color Constancy Project Color Constancy Project

github

Empirical study of colour constancy under variable illumination. Includes sensor calibration and controlled Google Pixel 3a experiments.

▸ Perspective Correction Project
Perspective Correction Project Perspective Correction Project Perspective Correction Project

github

Geometric rectification of architectural photos using camera calibration, vanishing-point estimation, and homography for sub-pixel façade alignment.

▸ Cat Mario AI
Cat Mario AI Cat Mario AI

github

NEAT-based RL agent that masters Cat Mario from raw video—drawing inspiration from MarI/O and CrAIg.

▸ VIM-RL

Multi-agent RL framework that guides a generic driving policy with specialist experts, producing 44% safer trajectories in occluded-pedestrian scenarios.

▸ LLaMA Profiler
LLaMA Profiler

github

“LLaMA-3.2 × Perfetto” traces every compute and memory step of a single transformer block—dot-products, RoPE, KV-cache I/O, MLP, layer norms, residual adds—and visualises them in Perfetto with rich tensor metadata. Quick start: run `setup.sh`, download weights via HF-CLI, then `torchrun load_model.py` to emit a Perfetto JSON trace.

▸ WaveFormer

paper

WaveFormer leverages discrete wavelet transforms to capture multi-scale token interactions in hierarchical transformers. It tops the leaderboard on FLARE, KiTS, and AMOS while cutting FLOPs.