Projects
Machine learning, deep learning, and time series forecasting — from thesis-level research to hands-on Kaggle competitions.
Predicts regional traffic density from sensor data across 21 districts of Madrid at 15-minute intervals. Uses a Bi-LSTM encoder-decoder architecture with CNN layers for sequence-to-sequence prediction (24-step lookback, 4-step forecast). Spatial correlation analysis significantly improved accuracy.
Predicts customer asset amounts using bank data warehouse, KKB credit bureau, and neighbourhood indicators — enabling targeted marketing campaigns. XGBoost selected as optimal model with MAE 98.6 and MAPE 17% on test data.
Full PyTorch implementation of the classic LeNet-5 — two convolutional and average-pooling layers, a flattening layer, two fully-connected layers, and a softmax classifier.
Implements and compares five kernel schema for face recognition classification, examining performance trade-offs across kernel choices.
Comprehensive ML pipeline implementing KNN, Decision Trees, Random Forest, HistGradient Boosting, and XGBoost with tuning via Grid Search, Optuna, Bayes Search, and Halving variants.
Probabilistic regression with Gaussian Processes using GPy — exploring kernel selection and uncertainty quantification for Bayesian non-parametric modelling.
Estimates hourly traffic intensity for the most congested Madrid district. Compares SARIMA and ETS statistical models against ML approaches on sensor data.
Analysis and forecasting of Turkey's annual export data using ETS decomposition and smoothing models.
Unsupervised NLP using LDA and NMF to extract and predict latent topics from text corpora.