From a first Titanic submission to a deployed fraud detector. Filter by what kind of work it is.
Phase 7 capstone on 284,807 highly imbalanced transactions (0.17% fraud). Weighted XGBoost beat Random Forest, ensembles, and a Keras NN on PR-AUC; deployed as a Streamlit app with CSV upload.
Phase 6/7 capstone on the Telco dataset. Compared class_weight vs SMOTE for imbalance, tuned via GridSearchCV for recall, and shipped a tuned Random Forest at 0.81 recall.
~5,572 texts, BoW vs TF-IDF across Naive Bayes and Logistic Regression. Naive Bayes + BoW won and got deployed; a from-scratch-embedding LSTM was built afterward for comparison and came in lower.
A Conv2D + MaxPooling2D + Dense network in Keras, 121,930 parameters, trained from scratch on the classic digit set.
Z-score outlier analysis on batting stats, plus a win-prediction model that found static pre-match features give near-random predictions — a real finding, not a failed model.
Time-series exploration of India's COVID case data with pandas and a set of Matplotlib/Seaborn dashboards.
Government agricultural data pulled via API, cleaned in pandas, and queried through a SQLite backend.
Kaggle competition — full EDA, feature engineering (titles, family size, age/fare bands), and a Random Forest pipeline.
Advanced regression on 79 features with log-transformed targets and a full preprocessing ColumnTransformer pipeline.
Kaggle's MNIST-as-CSV competition, solved with a Random Forest on raw pixel values before the later CNN revisit.
GridSearchCV combined with pipelines on the Breast Cancer dataset to systematically tune model performance.
Head-to-head comparison of five classification algorithms on the classic Iris dataset.
Linear regression model predicting farm yield from agricultural input data.
Real-time weather dashboard for Indian cities, built on the Open-Meteo API.