Portfolio
Here you can explore some of the projects I have worked on and the skills I have developed.
A distributed Apache PySpark pipeline that forecasts hourly electricity demand from time-series and weather data. A Random Forest outperformed the Linear Regression baseline (compared by RMSE).
A leakage-free classification pipeline on the classic Titanic passenger dataset, with feature engineering (title extraction, family grouping) and a Gradient Boosting model tuned with Bayesian search. 83% accuracy on unseen passengers (baseline 61.7%), explained with SHAP.
An end-to-end regression pipeline on 4,140 Seattle-area home sales, comparing Ridge, Random Forest, XGBoost and LightGBM with 5-fold cross-validation. The tuned LightGBM model cut the mean absolute error by 59% versus the baseline.
Groups 11,121 car models into natural market segments (Mainstream, Performance / Luxury and Older Models) with K-Means, choosing K by silhouette score. The model is served through a FastAPI endpoint.
An AI chat app built with Streamlit on the Groq LLM API that answers Python questions with clear explanations and code examples, using custom prompt engineering.
Classifies 14,640 airline tweets as negative, neutral or positive. Text is cleaned and turned into TF-IDF features, then five models are compared with cross-validation against a baseline; Logistic Regression won with 79% accuracy.
Exploratory data analysis of traffic accidents in Porto Alegre. Focused on identifying high-risk zones, temporal patterns, and key contributing factors to support data-driven urban safety decisions.
Predicts students' AI dependency level (Low / Medium / High) from survey data on 15,000 students. Six models are compared in a leakage-safe, cross-validated pipeline; the tuned Logistic Regression reached 0.66 F1, +0.40 over the baseline.
An executive dashboard for a fictional Brazilian PC-hardware retailer. Star-schema model with a Calendar table and DAX measures for revenue, margin, return rate and customer satisfaction, each tracked against a target.
Exploratory analysis of 218 Amazon best-selling Valentine gifts. Food & Sweets and Jewelry draw the most reviews, while a correlation of just 0.07 shows that higher ratings barely predict popularity.
A responsive book catalog with search, 8 categories, favorites and a light/dark theme. Books and their files are stored in Supabase, so users can add and delete titles from the browser.