GPT-2-style LLM built from scratch in C/CUDA with hand-written backprop, BPE tokenizer, FlashAttention, pretraining, and SFT.

📊 Compare few-shot text classification with DistilBERT and TF-IDF + SVM using IMDB data, analyzing performance across various sample sizes.

An AI-powered web application that generates intelligent text summaries from images using the BLIP deep learning model. Built with Python, Flask, and HuggingFace Transformers. Upload any image and get an instant AI-generated description.

Few-shot text classification baseline: DistilBERT fine-tuning vs TF-IDF+SVM on IMDB. Includes metrics.json, learning curves, and qualitative error examples

Emotion analysis of song lyrics using NLP (DistilBERT) to explore emotional patterns across lines and overall sentiment progression

Game Recommender

Description: BERT-powered NER for industrial maintenance logs with Streamlit UI, batch processing, and Dockerized deployment. Extracts Faults, Components, Actions, and Equipment from unstructured text.