Projects

VayuBuddy: An LLM-Powered Air Quality Assistant

📝 Read Paper • 🚀 Try the App

VayuBuddy is an open-source conversational agent designed to make air quality information accessible, explainable, and interactive. It is powered by a function-calling large language model that can query and summarize historical AQI data, generate interactive plots, offer pollution comparisons across locations, and explain the causes behind air pollution trends — all through natural language dialogue.

VayuBuddy integrates code generation, and multi-function calling over structured datasets and live APIs. It enables citizens, journalists, students, and researchers to ask detailed environmental queries without any programming knowledge. This system highlights the potential of LLMs in scientific reasoning, civic tech, and environmental transparency.

AQBench: A Benchmark for LLM Reasoning on Air Pollution Data

AQBench is the first benchmark specifically designed to evaluate large language models on structured, executable reasoning tasks using real-world air quality data. It contains over 10,000 queries across seven reasoning categories — from spatial aggregation and temporal trends to policy funding analysis — each paired with verified Python code answers.

The benchmark tests a model’s ability to generate executable and correct code that analyzes real datasets from the Central Pollution Control Board (CPCB), NCAP funding, and state-level demographics. Models are scored using exec@1 and pass@k metrics, with in-depth error analysis for syntax, column selection, and statistical reasoning.

AQBench aims to promote safe, trustworthy, and domain-aware LLMs for high-stakes applications like climate policy, public health, and environmental monitoring.

StoryForge AI: Multi‑Modal Story Generator

📂 GitHub

StoryForge AI is an end-to-end creative AI pipeline that transforms user input keywords into an immersive multimedia story. Developed as part of a deep learning coursework project, the system combines natural language generation, image synthesis, and speech generation to create complete AI-generated stories.

Key Workflow

  1. Prompt Input
    Users provide a set of keywords or a short theme that acts as the creative seed.

  2. Narrative Generation
    A large language model (LLM) constructs a multi-paragraph story based on the prompt, ensuring logical flow and engagement.

  3. Visual Generation
    Each sentence or scene in the story is converted into a visual using diffusion-based generative models (e.g., Stable Diffusion or DALL·E).

  4. Speech Synthesis
    The story text is converted into spoken narration using text-to-speech (TTS) models, adding a human-like audio experience.

  5. Media Assembly
    The visuals, audio, and text are stitched together into a short video using basic video editing pipelines, delivering a fully immersive storytelling experience.

This project showcases strong coordination across multiple AI modalities and demonstrates how generative AI can be applied in educational tools, interactive fiction, and creative media production.

Traffic Congestion Mapping Using Google Maps + Mapbox

This project visualizes real-time traffic congestion across Indian metro cities like Mumbai, Delhi, Bengaluru, and Ahmedabad using Google Maps and Mapbox APIs. The data is processed into density heatmaps and congestion-level plots to support urban planning and traffic optimization. It demonstrates how mobility data can be turned into actionable intelligence using spatial data pipelines and visualization tools.

ANPR (Automatic Number Plate Recognition)

Developed a real-time Automatic Number Plate Recognition system using OpenCV and deep learning. The pipeline detects and extracts number plates from vehicle images/videos and recognizes the alphanumeric text using OCR models. This system can be integrated into traffic monitoring, parking management, or toll automation systems. Techniques used include image pre-processing, contour filtering, YOLO-based detection, and Tesseract OCR.

Smoking Detection Using Computer Vision

Created a real-time smoking detection tool to enhance workplace and public area safety. The system uses deep learning and image classification to recognize smoking behavior in video frames. Deployed using OpenCV and a custom-trained CNN, it alerts when smoking activity is detected in restricted zones, showcasing the use of vision-based AI for behavior monitoring.

PixEnchar: AI-based Image Upscaling Tool

PixEnchar is an image enhancement tool built using Enhanced Super-Resolution GANs (ESRGAN) and generative models. It upscales low-resolution or blurred images while preserving fine details and color fidelity. The tool is designed for photographers, surveillance footage analysis, and restoring old/damaged images using AI-powered restoration.

Abusive Word Detection in Videos Using NLP

This project involves detecting abusive language in video content by converting audio to text and running it through NLP pipelines. Leveraging Hugging Face models for profanity detection, it provides time-stamped flags for moderation and censorship. The system enables automated monitoring of online content for community guideline violations.