Projects
YouTube ETL Pipeline and Streamlit Dashboard
Grand Forks, ND
- Developed a full-stack ETL pipeline to automate the extraction of public video data from the YouTube Data API, including metadata like title, views, likes, and publish dates.
- Implemented a modular architecture (extract, transform, load) using Python and pandas, saving data to a local SQLite database for storage and future analysis.
- Designed transformation functions to clean and enrich the dataset by converting formats, generating video age metrics, and removing duplicates.
Robust Reinforcement Learning for Autonomous UAV Landing
Grand Forks, ND
- Engineered a Deep Q-Network (DQN) control system in PyTorch to autonomously land a UAV, implementing the full agent lifecycle including an epsilon-greedy policy, experience replay, and a custom trainer to solve the Bellman equation for optimal throttle control.
- Designed and built a high-fidelity 2D aircraft simulation in Python using the Pymunk physics engine, modeling complex aerodynamics with force-lookup tables and architecting a state representation that incorporates temporal dynamics for improved agent decision-making.
- Developed a resilient learning framework by introducing a simulated communication loss module within the environment, successfully training the agent to maintain flight stability and achieve a 100% landing success rate despite receiving delayed state information.
Feed-Forward Neural Network for Aircraft Pitch Prediction
Grand Forks, ND
- Developed a feed-forward neural network (FNN) in Python to predict aircraft pitch from flight characteristics (5400+ data points), incorporating He initialization, gradient clipping, and early stopping (patience=20) for robust training.
- Optimized model performance through automated hyperparameter tuning (10+ randomized trials), identifying stochastic gradient descent with ReLU activation and a 0.05 learning rate as the superior configuration, achieving a minimum training MSE of 0.0350.
- Evaluated the optimized model's generalization on unseen test data, yielding a rescaled Test MSE of 2.5609, and implemented visualization tools (loss curves, actual vs. predicted plots) for comprehensive performance analysis and reporting.
Feedforward Neural Network for Car Price Prediction
Grand Forks, ND
- Developed a NumPy-based Feedforward Neural Network to predict used car prices, implementing a comprehensive preprocessing pipeline handling currency/unit conversions, text parsing, and engineering a 'Car Brand Region' feature from raw data (5500+ entries).
- Analyzed complex feature interactions and data distributions through 13+ visualizations (histograms, scatterplots, heatmap) and compared Batch vs. Stochastic Gradient Descent, demonstrating SGD's superior convergence via RMSE tracking over 100 epochs.
- Implemented feature normalization (z-score) and a prediction function capable of handling scaled/unscaled inputs, saving trained model parameters, and visualizing learning curves for performance evaluation and model persistence.
Modeling Multiple Relationships Utilizing Multiple Linear Regression with Gradient Descent
Grand Forks, ND
- Developed an MLR model optimized with Gradient Descent, achieving an R² score of 0.683.
- Implemented data preprocessing (standardization) and hyperparameter tuning (learning rate=0.01) to enhance stability and convergence.
- Evaluated performance using MSE, MAE, R², SMAPE and visualized results with actual vs. predicted plots and 3D regression surfaces.
Training a Feedforward Neural Network for XOR Classification: A Comparative Study
Grand Forks, ND
- Designed and trained a feedforward neural network using stochastic and batch gradient descent, achieving a minimum MSE of 0.0073.
- Conducted comprehensive hyperparameter analysis across 12 configurations (varying learning rates, bias inclusion, training strategies).
- Visualized and evaluated convergence, demonstrating stochastic training with bias and a learning rate of 0.10 as the fastest and most accurate.
Pixel-Level Detection of AI-Generated Images Using CNNs and Grad-CAM
Grand Forks, ND
- Conducted an in-depth literature review and proposed a CNN-based pixel-level anomaly detection framework.
- Curated and partially preprocessed the ArtiFact dataset (31,933 images), targeting over 95% classification accuracy.
- Designed a supervised learning pipeline incorporating Grad-CAM visualizations for model interpretability and benchmarking.
Development of a Focused Web Search Engine with TF-IDF-Based Page Ranking
Grand Forks, ND
- Developed a focused web crawler extracting and ranking pages based on TF-IDF relevance, starting from seed URLs.
- Designed a Python back-end integrating content extraction, hyperlink crawling, and dynamic priority ranking.
- Engineered a user interface for queries and ranked results, optimizing for scalability to 100+ pages.
Focused Web Crawler and Search Engine with PHP and MySQL
Grand Forks, ND
- Designed and implemented a focused web crawler (using lynx, PHP, shell scripting) to retrieve and index up to 100 topic-relevant pages.
- Engineered a relational MySQL database to store crawled data, keywords, and relevance scores for efficient search.
- Developed an interactive PHP front-end for crawling, indexing, searching, resetting datasets, and visualizing ranked results.
Labeling & Classification of Income Tax Fraud (NY State Dataset) using Python
Grand Forks, ND
- Detected and classified fraudulent records with 99% accuracy using K-Means clustering and Isolation Forest anomaly detection.
- Engineered and assessed three classification models, selecting Extra Trees as optimal (99% accuracy, 1.09s compute time).
Graph Search Algorithms: DFS & BFS for Maze Exploration
Grand Forks, ND
- Deployed DFS and BFS algorithms in Python, achieving a 15% improvement in pathfinding efficiency over baseline.
- Established a systematic approach for node data management (using dictionaries and stacks), leading to a 40% improvement in system responsiveness.
Optimized Q-Learning for Continuous State Control in CartPole-v1
Grand Forks, ND
- Developed a Python Q-learning agent achieving an average reward of 195 after 500 training episodes.
- Engineered state-space discretization to adapt tabular Q-learning for continuous state variables.
- Improved agent performance by ~200% over baseline through strategic hyperparameter tuning (learning rate, discount factor, epsilon decay).
Database-Driven E-commerce Bookstore with User and Admin Functionalities
Grand Forks, ND
- Engineered a functional online bookstore using Oracle 21c supporting user accounts, browsing, purchasing for 50+ users, and admin book management (100+ titles).
- Implemented secure registration/login/logout and admin functionalities using HTML, CGI, Perl, Java, SQL.
- Developed dynamic title-based search and detailed book/customer info display with clickable hyperlinks, including a "Display Source" feature for 5000+ lines of code.
User-Friendly Mobile App for Browsing and Purchasing Books
Grand Forks, ND
- Developed a mobile commerce app for book browsing, purchasing, and user account management with secure login.
- Implemented dynamic price range searching (0−10, 10−20, etc.) via a streamlined mobile interface.
- Engineered administrator functions for data entry, system reset, and detailed book/customer views via hyperlinks.
Sales Focused Job Finding Web Application
Sfax, TUNISIA
- Designed and developed a full-stack web application using React.js (front-end) and Node.js/Express.js (back-end) following the MVC architecture, resulting in a scalable and maintainable codebase.
- Integrated Stripe.js for online payment processing, reducing transaction errors by 20%, and implemented JWT-based authentication to ensure secure user access.
- Optimized application data storage by utilizing MongoDB NoSQL SaaS solutions, improving database query performance by 25% and enhancing user experience.