Applied ML

A binary classification model predicting diabetic retinopathy probability from retinal image features. Uses backward elimination and AIC for variable selection on 1,151 patient records — providing a clinically interpretable probability score for early DR detection.

Clinically interpretable probability score for early diabetic retinopathy detection

R · Binary Logistic Regression · AIC

A binary classification system predicting 10-year coronary heart disease risk from patient vitals and lab results. Benchmarks Random Forests against Mixture Discriminant Analysis across 10 random seeds on 4,080 Framingham Heart Study patients for consistent, reliable predictions.

84.3% classification accuracy

R · MDA · Random Forest

A ConvLSTM deep learning model that combines spatial feature extraction with temporal sequence modeling for automated video classification. Processes raw video frames through an OpenCV pipeline and outputs class predictions for video-heavy operational workflows.

84% accuracy and 97% AUC

Python · TensorFlow · OpenCV

A full Transformer architecture built from scratch in PyTorch — no Hugging Face, no abstraction shortcuts. Covers positional encoding, multi-head self-attention, feedforward layers, encoder and decoder blocks, and the complete encoder–decoder stack with configurable depth and heads.

Full architecture from scratch — every component implemented in detail

Python · PyTorch · Lightning

A CNN+LSTM hybrid pipeline forecasting per-product daily purchases using lagged and calendar features. Predicts next-day demand for multiple products simultaneously in a single forward pass — enabling smarter inventory planning and procurement decisions.

Multi-output sequence model predicting demand for N products in a single forward pass

Python · PyTorch · pandas · scikit-learn

An unsupervised market basket analysis system applying the Apriori algorithm to uncover hidden product co-purchasing patterns from transaction data. Generates rules ranked by support, confidence, and lift — surfacing actionable bundling and cross-sell opportunities.

Actionable co-purchase patterns with high-lift rules for the retail domain

Python · pandas · mlxtend · matplotlib

A deep neural network pipeline predicting insurance policy purchase likelihood from customer data. Combines scikit-learn preprocessing with TensorFlow model training to output purchase probability scores for targeted marketing and personalized policy recommendations.

70% accuracy and 85% ROC-AUC, ahead of classical baselines

Python · TensorFlow · scikit-learn · pandas

A lightweight scikit-learn pipeline predicting insurance purchase behavior using Logistic Regression, Random Forests, and SVMs. Built as the classical ML counterpart to PolicyPredict — directly comparing interpretability and performance tradeoffs against the deep learning version.

Side-by-side deep vs. classical ML comparison on the same insurance dataset

Python · scikit-learn · pandas · NumPy

Association rule mining on 41,188 Portuguese bank clients to identify which attributes drive term deposit subscriptions. Rules ranked by standardized lift surface the most actionable client segments for targeted campaign outreach and rejection avoidance.

13 rules, with marital status and prior contact the top subscription drivers

R · arules · Apriori Algorithm

A non-parametric spline regression system modeling China's GDP growth from 1960 to 2014. Compares standard cubic splines across varying degrees of freedom against smoothing splines across bandwidth parameters — analyzing the bias-variance tradeoff for complex economic trend forecasting.

Smoothing splines outperformed cubic splines for exponential growth modeling

R · Cubic Splines · Smoothing Splines

A multiclass grain classification system built entirely in R, comparing five algorithms — Bagging, Random Forests, Boosting, MDA, and ANNs — on geometric and shape features. Automates grain type identification to reduce manual inspection effort in agricultural quality control.

Random Forest and Boosting consistently outperformed baseline classifiers across all grain types

R · scikit-learn · Boosting

A regression benchmark comparing Least-Squares, Lasso, and Ridge on the Boston Housing dataset across 506 properties and 13 predictors. Demonstrates regularization benefits in the presence of multicollinearity — with Lasso producing the most parsimonious model by excluding INDUS and AGE.

LASSO performed best, dropping INDUS and AGE for the most parsimonious model

R · glmnet · Cross-Validation

A non-linear regression benchmark comparing GAMs, MARS, and Ridge on 414 Taipei City properties. Exploits known non-linearities — house age takes a parabolic shape, MRT distance is inversely proportional to price — to demonstrate where non-parametric methods outperform linear regularization.

GAMs performed best, with MRT distance the dominant price predictor

R · earth · gam · ridge

A tree ensemble regression benchmark predicting compressive strength from compositional features — comparing Bagging, Random Forests, and Gradient Boosting on 1,030 specimens. Demonstrates ensemble method tradeoffs in low-sample, high-feature-interaction regression settings.

Boosting significantly outperformed Random Forest and Bagging algorithms

R · randomForest · gbm

An unsupervised clustering system classifying forensic glass samples from crime scenes by chemical composition. Compares Hierarchical Clustering, K-Means, and Gaussian Mixture Models on 214 samples across 9 chemical variables to support forensic evidence identification.

K-Means performed best, separating headlamp glass from all other types

R · mclust · K-Means

A non-parametric regression system modeling Russian life expectancy trends from 1800 to 2016. Compares k-NN regression and Gaussian kernel smoothing across a range of hyperparameters — analyzing the bias-variance tradeoff for complex non-monotonic time-series data.

Kernel smoothing preferred for smoother boundary behavior

R · k-NN Regression · Kernel Smoothing

A comparative regression system predicting vehicle fuel economy from 7 characteristics on 392 vehicles. Benchmarks LASSO against MARS — demonstrating MARS's advantage in handling non-linearity and variable interactions while both address multicollinearity in the Auto MPG dataset.

MARS significantly outperformed LASSO

R · glmnet · earth

A model benchmarking framework comparing seven tree-based classifiers — Decision Tree, Random Forest, AdaBoost, Gradient Boosting, LightGBM, XGBoost, and CatBoost — under automated PSO-based hyperparameter search. All experiments tracked via MLflow for reproducible comparison and downstream deployment.

PSO-tuned models beat default settings, with every run logged in MLflow

Python · scikit-learn · MLflow · hyperparameter-optimizer