💻 My Projects
PaliGemma
End-to-end reconstruction of PaliGemma in PyTorch, featuring SigLIP ViT, Gemma decoder with RoPE & GQA, KV caching, and autoregressive multimodal inference.
Pragna-1B
Reverse-engineering Pragna-1B. A from-scratch PyTorch runtime with RoPE, GQA, RMSNorm, SwiGLU, and checkpoint parity.
Wav2Vec2
End-to-end PyTorch implementation of Wav2Vec2 with convolutional feature encoding, Transformer contextualization, Gumbel vector quantization, and full pipelines for contrastive pre-training and CTC fine-tuning.
VisionSmith
A clean, from-scratch Vision Transformer (ViT) implementation in PyTorch, trained on MNIST based on the "An Image is Worth 16x16 Words" paper.
ArXivGPT
A 124M-parameter GPT-2 model trained from scratch on ArXiv abstracts, featuring fully custom Transformer components implemented in PyTorch.
ScratchSpeakSLM
A tiny GPT-style language model built completely from scratch in PyTorch, inspired by TinyStories, demonstrating how small transformer models can learn to generate coherent text.
LlamaCare
A LoRA-fine-tuned TinyLlama-1.1B model designed for supportive, empathetic mental-health responses. Trained on a custom 15k-sample dataset with 4-bit quantization, achieving massive perplexity improvements and efficient performance on a single H100 GPU.
Conformer
End-to-end Conformer ASR in PyTorch featuring Macaron-style blocks, relative multi-head attention, depthwise convolutions, and an LSTM decoder with CTC loss.
DeepSpeech2
A from-scratch PyTorch implementation of the DeepSpeech2 architecture for automatic speech recognition (ASR)
Indic-ASR
Hindi Speech Recognition model reducing WER from 30.18% to 20.75%, matching performance of commercial APIs like Sarvam.
CerinaFoundry
specialized multi-agent AI system designed to assist mental health professionals in drafting Cognitive Behavioral Therapy (CBT) clinical protocols.
Agentorithm
A sophisticated, multi-agent AI system designed to solve Data Structures and Algorithms (DSA) problems by breaking them down into simpler steps.
ArXivMind
ArXivMind is an intelligent, multi-agent system designed to accelerate academic research. Built with Microsoft's AutoGen framework, it automates the process of finding and summarizing scholarly articles from arXiv.
CFAspire
CFAspire is an intelligent chatbot designed to help users study for the CFA Level 1 exam.
StockSage
A powerful, multi-agent financial analysis tool built with the phidata library for deep diving into stock market data.