River Samet

I built QTRAP (Quantum mechanics/molecular mechanics Trajectory Running, Analysis, and Preprocessing), the first publicly available end-to-end quasiclassical quantum mechanics/molecular mechanics dynamics workflow used to analyze femtosecond-scale dynamic effects within enzymes. The full pipeline is included here with Python and Bash source code, worked examples, and documentation.

• Includes permission-scoped tools and documentation that let an AI agent run the full workflow on a remote HPC cluster with minimal input
• Reduces hands-on time required to preprocess, run, and analyze simulations from weeks to just hours

Skills applied

• Python
• Bash
• agentic workflows
• HPC/Linux
• workflow design

I used custom convolutional neural network architectures (with/without data augmentation) and MobileNetV2 transfer learning models to classify 6,164 Chinese traffic sign images (58 classes). I packaged the best-performing model as a command-line tool that reports predictions and confidence for user-supplied images.

• Fine-tuned MobileNetV2 reached ~98% test accuracy vs. 96% for the best custom CNN
• Confusion matrices showed errors spread thinly across classes rather than concentrated in confusable signs

Skills applied

• TensorFlow
• Python
• CNNs
• transfer learning
• data augmentation

I loaded 120 years of Olympic results (Kaggle) into a Neon PostgreSQL database, wrote the SQL to reshape the data for overarching questions I had about the dataset, and built interactive Tableau dashboards to present what I found. Key findings are annotated on the dashboards; the README covers the rest of the analysis.

• Reshaped the data with CTEs, window functions, and multi-table joins to compare countries, sports, and eras
• Tableau dashboards use interactive filters and navigation, with findings written directly onto each dashboard

Skills applied

• Tableau
• SQL
• PostgreSQL
• dashboard design
• SQLAlchemy

I predicted electrocatalyst performance for the hydrogen evolution reaction from experimental data (UCLA) using random forests for classification and multi-layer perceptrons (MLPs) for regression, helping to guide future experiments to optimize catalytic activity for this reaction.

• Random forest classification reached a test F1 of 0.76 (0.78 on an alternative split); tuning closed the train/val F1 gap from 0.31 to 0.09, minimizing overfitting
• MLP regression reached a test RMSE of 0.11, with errors consistent across all three splits

Skills applied

• scikit-learn
• Python
• random forests
• MLPs
• permutation feature importance

I cleaned and reshaped a cardiovascular-health dataset (freeCodeCamp) and built categorical plots and a correlation heatmap with pandas, Matplotlib, and seaborn to determine which measurements track with disease.

• No single feature correlated strongly with cardiovascular disease presence; the strongest sat between 0.17 and 0.33
• Systolic blood pressure, cholesterol, age, and weight showed the strongest associations with cardiovascular disease presence

Skills applied

• Python
• pandas
• seaborn
• data cleaning
• correlation analysis