River Samet

Models electrocatalytic activity for the hydrogen evolution reaction on 545 experimental samples with 13 features, collected by a UCLA chemistry group. Random forests classify catalysts as good or bad; multilayer perceptrons predict overpotential directly. The comparison is the point — two model families, the same data, with cross-validation, hyperparameter tuning, and permutation feature importance used to interrogate both.

• Random forest reached a test F1 of 0.76 after tuning, up from a baseline that scored a perfect 1.00 on training data and 0.69 on validation — textbook overfitting.
• Plotting a representative decision tree showed why: leaves were splitting on a handful of points. That diagnosis drove the depth and leaf-size tuning.
• MLP regression reached a test RMSE of 0.11, with errors consistent across train, validation, and test.
• Co, Ni, Se, V, and reaction time were the most influential features — but the ranking shifted with architecture and scaling, so it's model-dependent rather than a universal chemical truth.

Skills applied

• scikit-learn
• random forests
• MLPs
• 5-fold CV
• hyperparameter tuning
• permutation importance
• overfitting diagnosis
• pandas

Four convolutional architectures compared on 6,164 Chinese traffic sign images across 58 classes from Kaggle: a baseline CNN, the same model with augmented training data, and two MobileNetV2 transfer-learning variants with and without trainable base layers. The best model was then shipped as a command-line predictor with a Bash wrapper that returns a labelled image and a confidence level.

• Both MobileNetV2 variants hit 98% test accuracy; the baseline CNN reached 96%.
• Data augmentation dropped accuracy to 68% — the baseline was barely overfitting, so augmentation mostly added noise to an already clean signal. A useful reminder that standard fixes need a problem to fix.
• Confusion matrices showed near-perfect diagonals, meaning errors were spread thinly across classes rather than concentrated in a few confusable signs.
• Stratified splits preserved the class imbalance, since some sign types are much more common than others.

Skills applied

• TensorFlow
• CNNs
• transfer learning
• data augmentation
• early stopping
• stratified splits
• confusion matrices
• CLI tooling

A full pipeline from raw Kaggle data to published dashboards: 271,116 athlete-event records from 1896 to 2016 loaded into a Neon PostgreSQL database with Python and SQLAlchemy, queried to build purpose-made tables, then visualized in interactive Tableau Public dashboards with written analysis attached to each view.

• Four framing questions drove the queries: how the Games expanded, how gender distribution shifted, how national medal counts compare, and how the first five Games differ from the most recent five.
• Queries range from grouped aggregates to window functions, CTEs, subqueries, and CASE statements — cumulative national medal totals didn't exist in the source data and had to be constructed.
• Dashboards use filters, calculated fields, parameters, and navigation buttons so a reader can move between the four questions without leaving the workbook.

Skills applied

• SQL
• PostgreSQL
• window functions
• CTEs
• SQLAlchemy
• Tableau
• dashboard design
• cloud databases

Exploratory analysis of patient medical data looking for relationships between physiological and behavioral features and the presence of cardiovascular disease. Covers feature engineering, outlier removal, reshaping with pandas, and two complementary views: a categorical plot of risk levels and a correlation heatmap.

• After removing implausible records, no single feature correlated strongly with disease presence — the strongest sat between 0.17 and 0.33.
• Systolic blood pressure, cholesterol, age, and weight led that ranking; smoking, alcohol use, and activity level showed little to none.
• The two views disagreed, and the write-up works out why: the categorical plot excludes continuous variables and was built before outlier removal.
• Correlation this weak doesn't support causal claims about a condition with this many interacting causes, and the analysis says so instead of overreaching.

Skills applied

• pandas
• seaborn
• feature engineering
• data cleaning
• outlier handling
• correlation analysis
• matplotlib