Building intelligence from first principles
I'm studying Robotics & Artificial Intelligence at Thapar Institute of Engineering and Technology. I believe in understanding things deeply, which is why I build everything from scratch.
After spending months exploring deep learning from the ground up, building GPT-2, diffusion models, autograd engines, and more, my current focus is on making real-world solutions with AI.
Currently serving as General Secretary of the Mecivions Research Society at Thapar, where I help foster a community of researchers and builders.
I'm actively looking for internships and opportunities, let's build something together.
A complete Stable Diffusion pipeline implemented from scratch, U-Net, CLIP text encoder, VAE decoder, and DDPM noise scheduler. Built entirely in PyTorch to understand every component of modern image generation.
Trained GPT-2 from scratch using PyTorch with bfloat16 precision and flash attention for efficient compute. Leveraged torch.compile and extensive hyperparameter tuning. Competes with OpenAI's released GPT-2.
Trained a toy decoder-only transformer on the Bhagavad Gita. A focused experiment in understanding how transformers learn to model language on a small, meaningful corpus.
A fully custom automatic differentiation engine built from scratch. Implements reverse-mode differentiation supporting core tensor operations and neural network primitives.
Trained an AI agent to play Flappy Bird using Deep Q-Network (DQN). Experience replay, epsilon-greedy exploration, and frame preprocessing for temporal awareness.
The University of Queensland · Brisbane, Australia
Conducted research on deepfake detection and explanation for social media video content. Developed AI explainability techniques and interpretable ML models for deepfake classification.
Mecivions Research Society · Thapar Institute
Leading the research society, fostering a community of AI/ML researchers and builders at Thapar.
I'm actively looking for internships and opportunities. Open to discussing collaborations, research, or anything AI. Let's connect.