projects

A multimodal RAG system that extracts and links information from text and images into a knowledge graph. This includes LlamaIndex + Hugging Face integrations and a Streamlit UI for interactive querying and visualization.

This LLM-based system, inspired by GAN design, uses a generator and two critics to generate and iteratively improve written code based on critics' feedback. The generator produces code while the critics evaluate its correctness and efficiency and provide feedback for further refinement.

This Python library streamlines data processing, model training and analysis in machine learning, with CLI commands and Python interfaces for dataset quality checks, format conversion, metric computation, plotting and model training.

An AI agent that connects to real-time databases through the MCP allowing natural language to SQL query translation and secure live data access. It supports full CRUD operations, enabling data management through natural language commands.

A web API that fetches and displays the latest research papers from arXiv based on user-specified topics and keywords, helping researchers stay updated with recent publications in their field.

A comprehensive library for interpreting and analyzing transformer language models. Built for researchers and developers who want to understand how LLMs work under the hood. Currently work in progress (hopefully will be better than TransformerLens ;)).

This tool scores your resume against a job description out of 100 and generates a detailed feedback highlighting your fit and suitability for the role. It is currently being extended to support multilingual resumes, thanks to a grant from Cohere Labs.

A prototype that evaluates the financial compliance of a transaction against a set of rules stored in Neo4j database using semantic search + rule-based system.

A PEFT-optimized chatbot powered by Llama-3.3-70B that translates user prompts into SQL queries efficiently. The model was fine-tuned on a custom dataset of 6K SQL query-natural language pairs and achieved 94% query generation accuracy.

A RL-based trading agent for the BSE Sensex that executes buy/sell trades, achieving an average annualized return of 9%. This was trained on 15 years of historical Sensex data and was benchmarked against RL algorithms like Deep Q-network, A2C and A3C.