A custom AI-powered customer service chatbot for seamless integration with e-commerce websites, providing 24/7 support for customer inquiries. The chatbot is built around a Retrieval-Augmented Generation (RAG) pipeline, combining document retrieval and generative models to deliver highly relevant, context-aware answers.

Using Chromadb and Hugging Face encoders, I created a fast document retrieval system, ensuring the chatbot can access the most relevant content quickly and efficiently for each customer query. With just one click, the system can embed both website content and external documents, guaranteeing that responses are always accurate and up-to-date.

This solution aims to empower businesses to manage higher volumes of inquiries while improving customer satisfaction and conversion rates through faster, more accurate, and consistent support.

To make AI more accessible for the Deaf and Hard of Hearing (DHH) community, I participated in the Google - American Sign Language (ASL) Fingerspelling Recognition Competition, which aimed to develop an AI system capable of detecting and translating ASL fingerspelling into text. The dataset provided by Google consisted of over three million fingerspelled characters recorded from 100+ Deaf signers under diverse lighting conditions, making it a challenging yet rewarding task.

I developed a Transformer-based sequence-to-sequence (Seq2Seq) model in PyTorch to accurately recognize and classify fingerspelled sequences from video frames. The model leveraged self-attention mechanisms to effectively track hand movements over time, ensuring robust predictions. To optimize real-time performance, I quantized the model and converted it into TensorFlow Lite (TFLite), making it deployable on edge devices like smartphones.

Through extensive data augmentation and preprocessing techniques, I improved the model’s generalization, reducing classification errors. My solution achieved a high accuracy score, demonstrating AI's potential in bridging communication gaps for the DHH community.

In this project, I explored the fascinating domain of Generative Adversarial Networks (GANs) by designing a model capable of transforming nighttime images into daytime images and vice versa. This problem posed a unique challenge due to the complex lighting variations and dynamic shadows that occur during the transition between night and day.

To create a robust dataset, I scraped and curated a collection of timelapse videos showcasing urban landscapes transitioning from night to day. Using this dataset, I trained a CycleGAN model with two generators: one responsible for transforming night images into day, and the other performing the inverse transformation. The generator architecture combined ResNet and U-Net, ensuring both high-quality texture preservation and spatial consistency. The discriminator was a CNN-based model, which refined the generated outputs by distinguishing real images from synthetically transformed ones.

The final model produced realistic image transformations with minimal artifacts, proving the power of unsupervised deep learning in image-to-image translation.

During my work in data science, I often encountered the need for customized web crawling and scraping solutions to collect structured and unstructured data efficiently. To address this, I developed ScrapDynamics, a powerful web crawling framework that enables users to explore, extract, and structure information from websites using regular expressions and automated data extraction techniques.

ScrapDynamics allows users to define custom crawling rules, extract links, metadata, and textual content, and handle large-scale web scraping tasks. It is optimized for high-performance data collection, with built-in support for asynchronous requests to maximize efficiency.

This tool has been invaluable in gathering datasets for various machine learning and NLP tasks, enabling automation in market research, competitor analysis, and information retrieval.

This project was centered around astrophysics and signal processing, specifically in detecting continuous gravitational waves (CW) from space. These signals are incredibly weak, requiring advanced techniques to distinguish them from background noise.

To tackle this challenge, I employed mel spectrograms, short-time Fourier transforms (STFT), and magnitude analysis to process raw wave signals. These transformed signals were then fed into deep convolutional neural networks (CNNs), trained to recognize the subtle patterns indicative of continuous gravitational waves.

The model showed promising results in identifying subtle wave patterns, offering a small step forward in the broader effort to study continuous gravitational waves.

A Portfolio Website built with React.js and Next.js, designed for speed, responsiveness, and a modern user experience.

Hosted on Vercel, it leverages client-side rendering (CSR) for optimized performance and features interactive project showcases, dark mode support, and dynamic content filtering.

This portfolio serves as a professional and easily maintainable platform to present my expertise in web development, data science, and AI.