Project 1: Web-Based EMS System

January 2024 – May 2024

Overview

A web-based EMS application developed for Sheboygan County, leveraging WIX/Google Sites for a user-friendly interface. This system ensures easy protocol indexing, real-time updates, and a streamlined layout based on EMS staff feedback.

Architecture & Tools

• Technologies: WIX or Google Sites for UI, MongoDB for data storage, SQL-based ER diagrams.
• Meetings & Feedback: Multiple sessions with EMS staff to refine protocols and user flow.

Key Achievements

• Implemented a searchable index for protocols, ensuring faster retrieval of critical information.
• Adapted the UI to changing backend requirements, maintaining a user-friendly experience throughout.
• Delivered a functional, easily maintainable final product after multiple feedback rounds.

Challenges & Solutions

• Rapid UI Changes: Used flexible site-building tools (WIX/Google Sites) to incorporate new EMS protocols without breaking the design.
• Integration: Ensured MongoDB data could be linked to a visual front end, bridging no-code tools with a robust backend.

Lessons Learned & Future Enhancements

• Integrating a custom front end for deeper control over the design and data workflows.
• Adding real-time alerts or push notifications for protocol updates.
• Expanding to a Docker-based deployment for easier scaling and portability.

Project 2: Full-Stack Fitness Tracker

October 2024 – February 2025

Overview

A full-stack web application that helps users log workouts, track progress, and manage personal profiles in a secure, responsive environment. Built to encourage consistent exercise and provide clear insights into performance improvement over time.

Key Achievements

• Engineered a full-stack app with React, Node.js, Express, and MongoDB, enabling up to 100 workout logs per day.
• Developed a secure RESTful API with 99% reliability in local testing, cutting response times by 37% from initial prototypes.
• Optimized MongoDB schemas, reducing query execution times by 25% under simulated production loads.
• Integrated React Hook Form for advanced validation, lowering form submission errors by 25%.

Architecture & Tools

• Front End: React (UI), React Hook Form (validation), Bootstrap (responsive design).
• Back End: Node.js, Express, RESTful API for CRUD operations.
• Database: MongoDB for user profiles and workout logs.
• Version Control: Git & GitHub with separate deployments for front and back ends.

Challenges & Solutions

• Data Consistency: Schema validation and server-side checks ensured integrity across user interactions.
• Performance Bottlenecks: Efficient indexing and schema design cut query times under higher load.
• Validation & Error Handling: Combined React Hook Form and Express validation to prevent invalid inputs.

Live URL

Click to view project live!

Lessons Learned & Future Enhancements

• Implementing user authentication and social login to expand user accessibility.
• Adding data visualization (Chart.js or Recharts) for more engaging progress analytics.
• Exploring Docker-based containerization for simpler, scalable deployments.

View GitHub Repository for Full-Stack Fitness Tracker

Project 3: Fake News Detection System

November 2024 – December 2025

Overview

A machine learning model to classify news articles as real or fake, comparing Logistic Regression (85% accuracy) and a fine-tuned BERT model (84% ROC-AUC).

Key Achievements

• Improved classification ROC-AUC by 6.8% using advanced preprocessing techniques.
• Optimized PyTorch model training, boosting performance by 15% over baseline methods.

Technologies Used

• Machine Learning: Logistic Regression, BERT (Hugging Face Transformers)
• Libraries: PyTorch, TensorFlow, Scikit-Learn
• NLP Techniques: TF-IDF vectorization, tokenization, text preprocessing
• Data Visualization: Matplotlib, Seaborn (ROC Curves, Bar Charts)
• Tools: Jupyter Notebook, Git, Pandas

Model Performance Comparison

Model	Accuracy (%)	ROC-AUC (%)
Logistic Regression	85.0	88.0
BERT (Fine-Tuned)	65.0	84.0

Key Challenges & Solutions

• Data Limitation: Addressed small dataset size with TF-IDF and tokenization for better feature representation.
• Overfitting Prevention: Applied dropout & weight decay in BERT training to generalize well.
• Performance Bottleneck: Leveraged PyTorch’s Trainer API for GPU-based training, boosting performance by 15%.

How It Works

1. Text Preprocessing: Tokenization, stopword removal, TF-IDF transformation.
2. Feature Engineering: Created ML-ready representations with TF-IDF & BERT embeddings.
3. Model Training: Logistic Regression vs. Fine-Tuned BERT using PyTorch.
4. Evaluation & Visualization: Compared models via accuracy, ROC-AUC, classification reports.
5. Real-World Testing: Users can input any article to check its authenticity.

Future Improvements

• Larger Dataset: Expanding training data for better generalization.
• Metadata Features: Incorporating author credibility & source reputation.
• Web Interface: Deploying as a user-friendly web app for public use.

View GitHub Repository for Fake News Detection System