SELECTED WORK

Projects & Research

From CUDA and computer vision to cloud-native AI platforms and mobile applications.

/ MASTER'S THESIS

Master's Thesis

MASTER'S THESIS

Innovative Rehabilitation Approach for Upper Limb Neurologic Conditions Using Mixed-Reality Simulation and EEG/EMG Biofeedback

My research focused on using Augmented Reality (AR), Mixed Reality (MR), Electroencephalogram (EEG), and Electromyogram (EMG) to assist stroke patients with upper limb weakness. Part of the "Smart NeuroRehab Ecosystem," in collaboration with the University of Washington Bothell, University of Washington Seattle, and Harborview Medical Center. Developed an interactive mobile-based AR/MR rehabilitation environment to enhance patient engagement. EEG and EMG data were processed in MATLAB with CNN, FNN, RNN, SVM, and LSTM models in PyTorch and TensorFlow for motor task classification.

THESIS DEMOS

GAME FLOW

GAME FLOW 01

GAME FLOW 02

View Publication

/ COURSE PROJECTS

Course Projects

Enhanced Vocabulary Trees for Real-Time Object Recognition in Image and Video Streams

This project explores k-Means clustering trees in computer vision, building on Nister & Stewenius (CVPR 2006). It evaluates their impact on visual recognition accuracy, structure, and efficiency. Implementation includes ORB feature detection, k-means clustering, max voting, and score calculations, tested on public datasets. Future work includes larger datasets (COCO) and alternative feature detectors to improve accuracy, advancing research in the Bag of Words framework.

CONTRIBUTORSSugam Jaiswal Josiah Zacharias

GitHub

CUDA Accelerated K-Means Clustering

This project implements an efficient k-means clustering algorithm on CUDA, leveraging GPU parallel computing for large-scale data clustering. A novel data partitioning technique accelerates key computations, particularly distance computation. Experiments show significant speedups over CPU-based implementations, making this approach promising for large-scale applications.

CONTRIBUTORSSunjil Gahatraj

GitHub

Searching Kitty

A web-based straightforward Word Search game written in Python and JS. Using React as the frontend client and Flask as the backend server.

CONTRIBUTORSWarren Liu Shahruz Mannan Amalaye Oyake Alan Jiang Barack Liu

GitHub

INKredible

A SwiftUI iOS application for OCR(Optical Character Recognition) with the feature to download the translated file into pdf, word or text format. This app also has a donations page.

CONTRIBUTORSSunjil Gahatraj

GitHub

/ PROJECTS

Projects

Real-time Particle Simulation with CUDA

This project leverages CUDA to accelerate a particle system simulation, enabling real-time visualization of fluid dynamics or physics simulations. Particles are simulated with gravity, elasticity, and spread parameters, and visualized in 3D using Python and OpenGL. The system can handle a large number of particles and frames, and simulates complex behavior like waterfall effects, while providing real-time interaction and performance optimization.

GitHub

Real-time Audio Feature Extraction with CUDA for Speech Recognition

This project uses CUDA to accelerate the extraction of MFCCs (Mel-frequency cepstral coefficients) from live audio streams for real-time speech recognition. By leveraging GPU parallelism, it speeds up the processing, reducing latency compared to CPU methods. The system extracts features from audio, which are then used by a speech-to-text model for real-time transcription, making it suitable for applications like virtual assistants and transcription services.

GitHub

Real-time Traffic Analysis with CUDA Object Detection

This project uses CUDA for real-time traffic analysis, focusing on vehicle detection and speed estimation. The system processes a sample video input, detects vehicles, calculates speeds, and visualizes results in real time. It includes key visualizations to understand vehicle speed distributions, variations, and trends. By leveraging GPU acceleration, the system demonstrates an efficient and scalable approach to real-time traffic monitoring and analysis.

GitHub

Chatbot Response Acceleration with CUDA LLM Inference

This project accelerates chatbot response times by using CUDA with ONNX Runtime to run language model inference locally. It enables faster, real-time interactions for customer service and other applications by converting models like GPT-2 into an optimized ONNX format for GPU execution.

GitHub

ConvoSphere: 3D GPT Chat with Avatars

ConvoSphere is a Unity-based 3D multiplayer chat app where users interact using customizable avatars. Powered by GPT-3.5, avatars respond intelligently in real time with text-based replies and emotive animations. Built with Photon PUN, it offers immersive AI-driven social interactions in a virtual environment.

GitHub

2048 Game

A simple and clean desktop version of the classic 2048 puzzle game, built using Python and Tkinter.

GitHub

Cloud-Based Resume Intelligence Assistant

Built a GenAI-powered application that analyzes resumes to extract skills, identify gaps, and generate ATS-friendly improvement recommendations using open-source LLMs and semantic similarity techniques.

GitHub

AI Startup Intelligence Platform

Built an end-to-end AI startup intelligence platform using Python, FastAPI, React, PostgreSQL, SQLAlchemy, Sentence Transformers, and Google Gemini, enabling semantic search and Retrieval-Augmented Generation (RAG) over startup funding, hiring, product launches, and market intelligence data. Developed a modular data ingestion pipeline, vector embedding generation, REST APIs, analytics dashboards, and an AI-powered analyst interface for natural-language querying of startup events.

GitHub

Scholarship & Career Intelligence Platform

Cloud-native platform that aggregates scholarships, internships, fellowships, research opportunities, and jobs through streaming pipelines while providing AI-powered recommendation, semantic search, and RAG-based career guidance using GCP, Kubernetes, BigQuery, and MLOps.

GitHub

/ MISCELLANEOUS

Miscellaneous

POC on Youtube Data(Big Data Analysis)

Developed proof-of-concept proposal for analyzing large YouTube’s datasets using Hadoop, HDFS, Map-Reduce, Hive, Pig, and other technologies to curate top 10 lists across various categories.

Advanced AI Transformers for Computer Vision

This project, part of the "Advanced AI: Transformers for Computer Vision" LinkedIn Learning certification, leverages advanced AI transformer models for computer vision tasks. It focuses on implementing transformer-based architectures for image classification and segmentation tasks, enhancing performance in visual recognition. The project involved hands-on practice with the latest AI transformers, improving efficiency and accuracy in computer vision applications.

GitHub