Lists (1)
Stars
An AI-powered Fashion Assistant built using Retrieval-Augmented Generation (RAG). Combines semantic search, reranking, and LLM-based reasoning to deliver intelligent fashion recommendations, produc…
An empirical study of Transformer adaptation techniques. Pre-training from scratch (MLM), classic fine-tuning, and from-scratch implementations of PEFT methods (LoRA, Adapters). Tuning both encoder…
From-scratch NumPy implementations of core deep learning architectures (DNN, CNN, RNN). A journey into the first principles of AI.
End-to-end implementation of Reinforcement Learning with Human Feedback (RLHF) to align a GPT-2 model with human preferences — covering Supervised Fine-Tuning (SFT), Reward Modeling, and PPO-based …
A hands-on exploration of Retrieval-Augmented Generation's core components: semantic search, retriever evaluation, and context-augmented LLM prompting.
Building production-grade Retrieval-Augmented Generation (RAG) systems. Explore advanced chunking strategies, hybrid search with Weaviate, LLM task orchestration, and fine-grained generation control.
An implementation of a classic N-Gram Language Model from scratch using PyTorch. This project covers word embeddings, perplexity evaluation, and t-SNE visualization for text generation.
A from-scratch implementation of Word2Vec models (CBOW and Skip-gram) in PyTorch. This repository includes model training, word similarity search, and t-SNE visualizations of the learned embeddings.
Automated document classification system using PyTorch & TorchText. Loads and preprocesses news articles, trains a text classification model, visualizes embeddings, and predicts topics such as Worl…
An end-to-end pipeline for adapting FLAN-T5 for dialogue summarization, exploring the full spectrum of modern LLM tuning. Implements and compares Full Fine-Tuning, PEFT (LoRA), and Reinforcement Le…
U-Net-based semantic segmentation on images - pixel-level road and environment understanding.
A dynamic web application built with Streamlit, offering real-time face recognition and verification capabilities. It leverages a pre-trained FaceNet model to accurately perform 1:1 identity matchi…
An end-to-end, real-time wake word detection system built with a custom hybrid Conv1D + GRU deep learning model. Features include programmatic data synthesis, robust audio signal processing, and pr…
A full end-to-end Neural Machine Translation (NMT) system built using TensorFlow/Keras. This encoder-decoder model translates English to French using LSTM-based sequence-to-sequence architecture wi…
A hands-on repository demonstrating word embeddings, featuring Word2Vec training with Gensim and text classification using pre-trained GloVe vectors in PyTorch.
Hybrid RAG with three retrievers—Lexical (BM25), Semantic (embeddings), and Hybrid (Reciprocal Rank Fusion) — parallel retrieval, Llama 3 generation, and side-by-side evaluation with reproducible n…
A clean, modular implementation of a decoder-only Transformer (mini-GPT) from scratch in PyTorch for autoregressive text generation.
Transforming a general-purpose LLM into a specialized code generator by fine-tuning <1% of its parameters using Low-Rank Adaptation (LoRA).
An advanced, end-to-end pipeline for building production-ready conversational AI agents using OpenAI's GPT-4o. This project demonstrates a complete system with security guardrails, structured outpu…
An end-to-end project where I built a Transformer from scratch, fine-tuned DistilBERT to 94.8% accuracy, and deployed it as a live interactive demo on Hugging Face.
An end-to-end RAG system that grounds LLMs in factual reality, using semantic search on real-time news to provide verifiable, context-aware answers.
A hands-on journey from classic NLP foundations to modern Large Language Models (LLMs). This repo explores N-gram models, tokenization techniques, NLP data pipelines, and interactive LLM chatbots -…
A production-focused blueprint for the complete AI application lifecycle, from architecting RAG pipelines and autonomous agents to their quantitative evaluation.