Projects

Designing a full-stack, LLM-powered "digital twin" that learns an individual's writing style from their email history, then generates emails in their voice. Architecting a modular feature–training–inference pipeline with React frontend, TypeScript backend, and vector-based RAG retrieval to support future fine-tuning of personalized language models

Built a cattle breed classification model using ResNet-18, achieving 95% accuracy and strong precision/recall across 8 breeds. Addressed class imbalance in the skewed dataset by applying class-weighting (assigning higher weights to underrepresented classes). Trained the model over multiple epochs for generalization before fully fine-tuning, leading to robust performance despite data imbalance.

A project which involves my effort to understand the intuition/logic behind research papers as well as the maths involved in these papers followed by code implementations.