Implemented a Tensorflow version of the Molecular VAE paper for automated molecular design. Trained on a dataset of 250,000 molecular graphs from the ZINC database, this deep learning variational autoencoder learns to encode molecular structures into a continuous latent space, enabling the generation of novel drug-like molecules. The model facilitates faster drug development by allowing researchers to explore chemical space more efficiently and generate molecules with desired properties.

Encodes molecular structures into continuous latent representations for generation

Variational Autoencoder

Trained on 250,000 molecular graphs from the ZINC database

Large Scale Training

Generates novel drug-like molecules by sampling from latent space

Molecule Generation

Python

TensorFlow

NumPy

Molecular Variational Autoencoder

An AI-powered accessibility tool that generates real-time audio descriptions for videos, making visual content accessible to blind and visually impaired users. Built for the Llama 4 Hackathon, it uses multimodal AI to analyze video frames, transcribe speech with Whisper, and generate natural narration synchronized with playback.

Auto-detects Chinese, Japanese, Korean, Arabic, Spanish, French, and more

Multi-Language Support

Choose voice, adjust speech rate and pitch to preference

Voice Customization

Over 2.2 billion people worldwide have vision impairments, yet most video content lacks audio descriptions. This tool changes that using Llama 4's multimodal capabilities.

**Smart Video Chunking** — Splits videos into 9-second segments for efficient parallel processing with FFmpeg

**Llama 4 Multimodal Analysis** — Analyzes video frames and generates contextually-aware, natural descriptions optimized for accessibility

**Audio Transcription** — Integrates OpenAI Whisper to preserve dialogue context for better descriptions

**Synchronized Playback** — Text-to-speech narration plays in sync with video, with customizable voice settings

React 19

Vite

Tailwind CSS

Node.js

Express

Llama 4 API

OpenAI Whisper

FFmpeg

Video Description Generator

A full-stack web application similar to Google Maps, built specifically for Rhode Island. Connected a React.js frontend with a Java backend server to create an interactive mapping experience capable of pathfinding using Dijkstra's algorithm. Utilizes a SQL database containing over 1 million road segments to calculate optimal routes between any two points in the state. Features include real-time route visualization, turn-by-turn directions, and efficient database queries for fast response times.

Rhode Island Maps

Overview

Key Features

Pathfinding Algorithm

Large Scale Database

Interactive Map UI

Tech Stack

Project Info

Contribution