Shaghayegh Rouhi

About Me

I’m Shaghayegh Rouhi, but I usually go by Shay. I’m currently pursuing a Bachelor's degree in Data Science at the University of Nebraska-Lincoln, with minors in Statistics, Mathematics, Computer Science, and Economics. I’m passionate about applying machine learning techniques to solve complex problems, particularly in areas like data modeling and predictive analytics.
In addition to my academic work, I enjoy exploring web development as a way to bring my data science projects to life and create user-friendly solutions. I’m always excited to learn new technologies and enhance my skills to tackle real-world challenges.
As I work toward my degree, I’ve had the opportunity to gain hands-on experience through various projects and internships, where I’ve honed my problem-solving, programming, and analytical skills.
Feel free to explore my projects and reach out if you'd like to collaborate or learn more!

Projects

Built an AI-powered web app that matches resumes to job descriptions using NLP, semantic similarity, and key phrase extraction.
Provides real-time, actionable feedback on missing skills and areas for improvement using state-of-the-art transformer models including KeyBERT and FLAN-T5.
Implemented Streamlit UI and hosted the app on Hugging Face Spaces for easy access by users.
Integrated CI/CD pipelines with GitHub Actions to automate testing and enhance code reliability and development speed.

Check it out: GitHub Repo

Utilized SQL and Power BI to analyze customer churn behavior in an e-commerce platform, deriving strategic business decisions.
Identified key churn trends such as higher rates among mobile users and new customers, supporting targeted retention strategies.
Built an interactive Power BI dashboard to visualize churn patterns across tenure, delivery distance, and complaint frequency, enabling stakeholders to make informed decisions.

Check it out: GitHub Repo

This project applies machine learning to analyze whether smaller county assessor’s offices should rely on real estate data from larger counties or their own local data. The model uses features like square footage, number of bedrooms, and location to predict how long it takes for a property to sell.

Technologies Used:

Python, Scikit-learn, Pandas, NumPy
Data cleaning, feature engineering, and model evaluation (MAE, RMSE)

The project aims to assist county offices in making more informed decisions about data usage.

Check out here for more information: ML-RealEstateXCounty

This project demonstrates the process of training a FeedForward Neural Network (FNN) from scratch, without relying on deep learning libraries. The notebook walks through the mathematical principles of FNNs, implementing gradient descent and backpropagation for a neural network with a single hidden layer and adjustable nodes.

Technologies Used:

Python, NumPy, Matplotlib
Neural Networks, Backpropagation, Gradient Descent

Check out here for more information: ML - Training from Scratch

Experience

Nebraska Department of Transportation (NDOT), Lincoln, NE

Data Scientist Intern | May 2024 – Present

Develop a Convolutional Neural Network (CNN) in PyTorch to detect sidewalks from geospatial imagery, supporting improvements in pedestrian safety infrastructure.
Build a predictive statistical model using Multiple Linear Regression in R to forecast federal performance measures and improve mid-period reporting and target adjustment.
Reimplement a snowplow performance study using R, Python, and GIS to evaluate efficiency during storms and determine resource needs.
Create interactive Power BI and Tableau dashboards to help stakeholders track performance over time using data cleaning, analysis, and visualization skills.
Conduct a validation study comparing AI-powered traffic cameras and traditional ATR sensors against manual counts, using statistical analysis to support investment decisions.

Undergraduate Research - Department of Mathematics, UNL

Undergraduate Research | September 2024 – Present

Develop a nonlocal deep neural network (DNN) training method replacing classical derivatives with integral operators, improving model accuracy and stability for discontinuous ReLU activations.
Validate the proposed nonlocal DNN with multiple datasets and parameters in Python (NumPy, TensorFlow), demonstrating effectiveness of nonlocal methods.
Achieve high accuracy in image recognition tasks using nonlocal operators.
Supported by the UCARE program, resulting in a poster presentation at UNL Research Day.

Microeconomics Learning Assistant & Tutor, Lincoln, NE

Learning Assistant (LA) & Tutoring | January 2024 – May 2024

Held weekly review sessions and exam prep workshops for 20+ students, improving understanding of course material.
Collaborated with professors to review course progress, grading, and student performance, applying organizational and teamwork skills.
Assisted students via course platforms, email, and meetings to address concerns and improve engagement.

Research

In this research, I explore nonlocal approaches to training neural networks, specifically focusing on improving the training process by addressing limitations in conventional activation functions, such as the ReLU (Rectified Linear Unit). ReLU has been widely used due to its simplicity, but it poses issues like instability and dead neurons due to its discontinuity at zero. By introducing nonlocal derivatives, which allow for the interaction of data points over a range of values rather than relying solely on local gradients, my work aims to provide smoother, more stable training dynamics for deep learning models. This approach has the potential to improve the convergence rate and overall performance of neural networks, especially in applications involving complex datasets like image recognition.

You can view or download my research poster here. Use zoom controls in your browser to resize.

View Poster Download Poster