A from-scratch PyTorch implementation of core transformer components including self-attention, masked attention, and multi-head attention following the seminal Attention Is All You Need paper.
A nanoGPT-inspired character-level transformer model to study tokenization, autoregressive decoding, and small-data generative modeling behavior.
Comparative evaluation of zero-shot, one-shot, and few-shot prompting techniques on dialogue summarization, revealing one-shot prompting as the most efficient for lightweight FLAN-T5 models.
A Retrieval-Augmented Generation system using Gemini LLM, Chroma vector DB for embeddings with metadata, and LangGraph for iterative retrieval and improved answer grounding.
An end-to-end forecasting system employing Transformers, XGBoost, Random Forests, Prophet, and DoWhy causal inference to model large-scale energy consumption patterns.
A 98%-accuracy human activity recognition pipeline using classical ML and deep time-series architectures including Conv-LSTM for rich temporal modeling.
A production-ready Random Forest credit scoring system achieving ~90% accuracy, deployed via Gradio for interpretable, real-time financial risk assessment.
A Siamese neural network using contrastive loss to detect anomalous patient sensor patterns, achieving 98% detection accuracy after balancing and optimization.
A DCGAN-based data augmentation pipeline for dermatoscopic images, improving model accuracy by 15% and reducing overfitting through realistic synthetic image generation.
A Tableau dashboard that visualizes customer churn experiment outcomes, retention patterns, and demographic segmentation from A/B test results.
A Tableau analytics dashboard exploring Netflix content performance, genre-level trends, and global distribution insights.