1. User Spending Prediction – Conducted (EDA) to identify key spending patterns using pandas, NumPy, and Matplotlib. Built and compared regression, classification, and ensemble models in scikit-learn (Linear/Logistic Regression, Decision Tree, Random Forest, Gradient Boosting) and evaluated results using R?, MAE, MSE, and Accuracy metrics.

2. Opossum Clustering Analysis – Performed (EDA) and applied (PCA) using pandas, NumPy, and scikit-learn. Implemented and compared K-Means, DBSCAN, and Fuzzy C-Means — to group opossums. Evaluated clustering quality using Silhouette Score, Davies–Bouldin Index, and Calinski–Harabasz Score, visualizing results with Matplotlib and Seaborn

3. Name Matching with Fuzzy Logic – Performed EDA and developed a fuzzy logic–based system in Python to evaluate similarity between names under uncertainty. Utilized fuzzywuzzy, soundex, and Damerau–Levenshtein distance algorithms integrated scikit-fuzzy to compute composite similarity probabilities.

4. Heart Disease Prediction – Developed a single-layer Perceptron in Python using PyTorch and NumPy with StandardScaler preprocessing. Trained the model using Sigmoid activation and evaluated performance for Accuracy and Loss.

5. Image Classification with CNN – Developed a custom Convolutional Neural Network in PyTorch for multi-class image recognition. Implemented convolutional and pooling layers from scratch, used Adam optimizer and CrossEntropyLoss, and evaluated model performance using Accuracy and Loss, visualizing predictions and training progress.

6. Development of Neural Networks for NLP – Processed Ukrainian text data using spaCy for linguistic analysis, then generated BERT embeddings using Hugging Face Transformers and compared sentence similarity with cosine similarity. Additionally explored GloVe word vectors for semantic comparison between words.

7. House Prices Prediction (Kaggle) – Performed comprehensive EDA with missing-value imputation, outlier removal, and feature engineering (e.g., TotalSF, Age). Encoded categorical features use one-hot encoding, applied log transformation to normalize target distribution, and trained an XGBoost Regressor tuned for performance. Model evaluation was based on the Root Mean Squared Logarithmic Error (RMSLE) metric, achieving high accuracy on validation data.