Zachary's Resume (pdf)
Data Science Courses Completed
Welcome to my data science portfolio! I'm Zachary Raup, a data scientist with a strong foundation in physics and a passion for uncovering insights from complex datasets. I graduated Summa Cum Laude from Kutztown University with a B.S. in Physics, where I focused on data modeling in astrophysical systems—particularly exoplanets and binary stars. This research experience trained me to approach problems analytically, work with real-world uncertainty, and extract meaning from noisy data.
To strengthen my data science skillset, I earned certifications from DataCamp in Data Science, Data Analysis, Python, and SQL, and completed coursework in machine learning, data preprocessing, and visualization. I apply these skills using Python (pandas, scikit-learn, matplotlib, numpy) and SQL, with additional proficiency in Tableau and Power BI for data storytelling.
My work focuses on building interpretable, performance-driven models to support real-world decision-making. I enjoy collaborating across disciplines, turning messy data into actionable insight, and constantly learning new techniques to grow as a scientist and developer.
Thanks for visiting—feel free to explore my projects!
Below is a list of data and coding projects that I have completed. Use project title links to view more information about each project.
Developed a deep learning pipeline to classify chest X-rays as Normal or Pneumonia using an ensemble of pretrained CNNs (ResNet18, DenseNet121, EfficientNet-B0). Achieved a 91.2% test accuracy and an F1-score of 0.9332, with all models demonstrating high pneumonia recall, minimizing false negatives.
Key components:
- Implemented 3-fold stratified cross-validation and early stopping to ensure robust model performance
- Applied class-weighted loss functions to address class imbalance during training
- Integrated Grad-CAM visualizations to highlight key diagnostic regions and enhance interpretability for clinical use
This project showcases how deep learning and explainable AI can support radiologists by improving diagnostic accuracy and transparency in medical imaging.
Skills Applied: PyTorch, Convolutional Neural Networks (CNN), Computer Vision, Deep Learning, Medical Imaging, Scikit-Learn and more
Grad-CAM visualization for a test chest X-ray correctly classified as Pneumonia by the ResNet18 model. The highlighted activation regions (in red and yellow) suggest the model focuses on areas of increased radiographic opacity within the central and lower lung fields—features often indicative of pulmonary infection. This supports the model's decision-making process and provides interpretability in a clinical context.
Confusion matrix for the ResNet18 model (fold 0) evaluated on the test set. The model correctly identified 384 of 390 Pneumonia cases (high sensitivity) and 185 of 234 Normal cases, resulting in 6 false negatives and 49 false positives. This performance highlights the model’s strong bias toward minimizing missed Pneumonia diagnoses, a clinically preferred trade-off in high-stakes triage settings.
Accurate weekly sales predictions are crucial for large-scale retailers like Walmart to optimize inventory management, labor allocation, and supply chain planning. This project uses historical sales data (2010–2012) to build a machine learning pipeline that predicts weekly sales using a blend of store-level, temporal, and economic features. After thorough EDA, feature engineering, and model tuning, five regression algorithms were evaluated, with XGBoost and LightGBM demonstrating top performance.
Key goals included:
- Comparing linear vs. non-linear models (Ridge, Decision Tree, Random Forest, XGBoost, LightGBM),
- Enhancing model generalizability through 5-fold cross-validation,
- Improving interpretability via residual diagnostics and feature importance visualization.
Best Model: XGBoost with RMSE ≈ $61.4K and R² ≈ 0.988 on the test set
Notable Insight: Holiday weeks and store-specific trends were the strongest predictors of weekly sales variability
Skills Applied: Python (pandas, scikit-learn, xgboost, lightgbm), Regression Modeling, Cross-Validation, Residual Analysis, EDA, Feature Engineering, Data Cleaning, Visualization
This dual-panel plot compares predicted weekly sales to actual values for the two best-performing models. XGBoost (left) shows a strong fit with tightly clustered points near the diagonal, indicating accurate forecasts and low bias. LightGBM (right) also performs well but with slightly more spread, particularly at higher sales values.
This figure compares the top 10 features driving predictions in both models. XGBoost heavily prioritizes the Store variable, revealing strong store-specific effects. LightGBM distributes importance more evenly across predictors such as Temperature, Fuel_Temp_Interaction, and Economic_Stress, suggesting broader generalizability.
This project applies unsupervised machine learning techniques to uncover patterns in Spotify audio data and recommend musically similar songs. By using Non-negative Matrix Factorization (NMF) for dimensionality reduction and t-distributed Stochastic Neighbor Embedding (t-SNE) for visualization, the feature space of over 6,000 tracks was mapped into an interpretable 2D projection. Cosine similarity was then used to identify songs most similar to “Blinding Lights” by The Weekend. The final result is an insightful visual and analytical exploration of musical relationships based on audio characteristics..
Skills Applied: Unsupervised Learning, NMF, t-SNE, Cosine Similarity, Data Preprocessing, Python (scikit-learn, NumPy, pandas)
This 2D visualization presents a t-SNE projection of the song dataset, with each point representing a track. Colors indicate cosine similarity to “Blinding Lights” by The Weeknd—green represents higher similarity, red represents lower. The 10 most similar songs are marked with square outlines, clearly highlighting clusters of tracks that share sonic traits. This approach visually demonstrates how audio features can be leveraged to identify stylistic similarity between songs.
The table below lists the 10 songs most similar to “Blinding Lights”, based on cosine similarity of audio features. These tracks span multiple genres and artists, yet share core musical qualities such as synth-driven production, emotional tone, and modern pop energy. The diversity of artists—from Post Malone to K/DA—illustrates the reach of “Blinding Lights’” sonic profile.
Top 10 Similar Songs to: Blinding Lights - The Weekend
- Blinding Lights (with ROSALÍA) — Remix - The Weekend, ROSALÍA (Similarity: 0.9988)
- I am not a woman, I'm a god — Halsey (Similarity: 0.9971)
- Feels Great (feat. Fetty Wap & CVBZ) — Cheat Codes, Fetty Wap, CVBZ (Similarity: 0.9951)
- Enough Is Enough — Post Malone (Similarity: 0.9939)
- MI FAI IMPAZZIRE — BLANCO, Sfera Ebbasta (Similarity: 0.9936)
- Last Day Alive — The Chainsmokers, Florida Georgia Line (Similarity: 0.9933)
- POP/STARS — K/DA, Madison Beer, (G)I-DLE, Jaira Burns, League of Legends (Similarity: 0.9931)
- Love It If We Made It — The 1975 (Similarity: 0.9929)
- When I’m Alone — Post Malone (Similarity: 0.9928)
- Together — Sia (Similarity: 0.9928)
Developed a machine learning pipeline to detect machine failures in a manufacturing environment using sensor data. Feature engineering and class rebalancing were applied to improve signal extraction and model fairness. Compared Logistic Regression, Random Forest, and XGBoost, with the latter achieving 98.6% accuracy and perfect AUC on the test set.
Key components:
- Created engineered features including
Power,Temp_Delta,Speed_Torque_Ratio, andWear_per_Torque - Applied SMOTE to correct class imbalance (only 339 failure cases among 10,000+ observations)
- Used SHAP for global and local model interpretability
- Achieved a strong balance of precision and recall, reducing false positives and false negatives
This project demonstrates how domain knowledge, feature construction, and ensemble methods can work together to create a reliable and explainable predictive maintenance solution.
Skills Applied: Machine Learning, Feature Engineering, SMOTE, SHAP, Ensemble Modeling, Supervised Learning, Python (scikit-learn, XGBoost, imbalanced-learn)
SHAP summary plot showing how each feature influences the model’s predictions. Tool_wear_[min], Power_[W], and Rotational_speed_[rpm] are among the most influential inputs. This visualization aids in validating model behavior and supports communication with non-technical stakeholders.
This confusion matrix shows the performance of the XGBoost model in detecting machine failures. It achieved 98.6% accuracy, with only 55 total misclassifications out of 3,865 samples. The model demonstrates excellent precision and recall, making it highly effective for predictive maintenance tasks in manufacturing.
This research project focuses on modeling the transit of exoplanets across stars using the Python package 'batman'. The objective was to accurately predict changes in stellar brightness during these transits, validated against photometry data from the CR Chambliss Astronomical Observatory (CRCAO). Methodologically, a physics-based model was developed and evaluated using a log likelihood function to fit observational data. The Markov Chain Monte Carlo (MCMC) algorithm, facilitated by 'emcee', enabled exploration of parameter uncertainties such as planet radius and transit timing. Visualizations created with matplotlib included light curves and histograms of parameter distributions. Presenting findings at the 241st AAS meeting highlighted contributions to understanding exoplanet transit dynamics, crucial for advancing knowledge of planetary systems beyond our solar system.
Light curve of TOI-4153 data (CRCAO) taken in a Blue (B) and Infrared (I) filter. The model is built using the Python transit modeler package 'batman'. The parameters of the model were determined using the Markov Chain Monte Carlo algorithm and known parameters taken from the ExoFOP database.
The goal of this project is to utilize MySQL queries to perform analysis of trends and relationships embedded within the Dognition database. Developed as a fundamental component of the ‘Managing Big Data with MySQL’ course from Duke University, the project focuses on refining and applying skills in data cleaning, sorting, and employing advanced analytical techniques using SQL. By exploring large datasets such as the Dognition database, the project aims to uncover meaningful insights into canine behavior patterns and preferences, leveraging robust data management practices to extract actionable intelligence for further research and practical applications in understanding and enhancing dog-human interactions.
This MySQL query analyzes geographical user trends within the Dognition dataset, resulting in a table that lists the top 5 states with the highest number of Dognition users.
Zachary's Portfolio
Project 1: Chest X-Ray Pneumonia Detection with Deep Learning
Project 2: Forecasting Retail Sales with Machine Learning | Regression Modeling
Project 3: Discovering Similar Songs using Machine Learning and Spotify
Project 4: Predictive Maintenance in Manufacturing
Project 5: Utilizing MCMC in Python to Explore the Parameter Space of an Exoplanet Transit
Project 6: Insights into Dog Behavior: Analyzing Dognition Data with MySQL