Project Archive

🎯 Context & Problem

Diagnosing Autism Spectrum Disorder in adults can be a lengthy and costly process. This project sought to answer: Can we use data from a standard psychometric screening questionnaire (the AQ-10) to build a Machine Learning model that identifies subtle patterns and offers an automated, rapid, and scalable initial risk assessment?

⚙️ Technical Methodology

The project began with unsupervised clustering (K-Means, DBSCAN) to understand the data's natural structure. The EDA revealed a significant class imbalance, which was addressed using techniques like SMOTE. A "bake-off" of multiple models with hyperparameter optimization was conducted, achieving an average F1-score of 0.87.

🛠️ Tech Stack

📊 Impact & Results

A validated model that serves as a non-invasive support tool for healthcare professionals, helping to prioritize cases and allocate resources more efficiently. It demonstrates the potential of AI to create more accessible, data-driven mental health solutions.

🎯 Context & Problem

A geriatric care company needed to validate a clinical hypothesis: is there a predictable relationship between hypertension peaks and the onset of psychopathological crises? The challenge was to structure a complex and sensitive data project rigorously, ethically, and in a way that generates preventive business value.

⚙️ CRISP-DM Methodology

A detailed plan including a sophisticated data preparation strategy with innovative variables like a 'Hydration' index. It proposed interpretable models (Decision Trees, Logistic Regression) to facilitate clinical validation and included a dedicated phase for the proactive identification of biases in historical data.

🛠️ Tech Stack

📊 Impact & Results

A strategic plan that ensures a solid foundation, minimizing risks and maximizing impact. It demonstrates high-level technical leadership and business vision. The blueprint is ready for execution by a development team.

🎯 Context & Problem

Beyond predicting individual diabetes cases, it was crucial to understand if natural groups of individuals with combined risk profiles existed. The objective: to shift from individual analysis to population segmentation to inform personalized prevention campaigns.

⚙️ Technical Methodology

An EDA → Feature Engineering → Clustering workflow. Key finding: "outliers" in health variables were a signal of high-risk groups, leading to the use of RobustScaler. Created composite variables: a healthy_habits_score and a cardio_risk_index. The Elbow Method justified 3 clusters as the optimal segmentation.

🛠️ Tech Stack

📊 Impact & Results

A strategic data asset: a segmented dataset that allows business analysts to explore risk profiles without ML knowledge. It democratizes insight for multidisciplinary teams.

🎯 Context & Problem

An international expansion decision represents a high financial risk. This project sought to replace intuition with a rigorous data science approach to answer: "Which countries offer the best balance of economic opportunity and stability for expansion?"

⚙️ Technical Methodology

A two-phase project: I) ETL & Preparation; II) Modeling & Segmentation. PCA was used to visualize relationships between countries. K-Means and DBSCAN were applied to group countries with similar profiles. Comparing methods provided a robust understanding of the data structure, resulting in a logical segmentation of nations.

🛠️ Tech Stack

📊 Impact & Results

A ranked and segmented list of candidate countries, providing a tool for evidence-based strategic decision-making. It reduces risk and optimizes investment in the global expansion strategy.

🎯 Context & Problem

The retro video game market has a fascinating economic behavior, but price data is scattered. The problem was: how can we systematically create a structured dataset from a dynamic web source efficiently and respectfully?

👤 My Role & Responsibilities

I acted as the sole Data Engineer, responsible for the entire process: research and feasibility (analyzing `robots.txt`), developing the Python scraper to handle dynamic content, implementing ethical practices (rate limiting), data cleaning and structuring, and the final dataset publication.

⚙️ Technical Methodology

The main challenge was the "infinite scroll". Instead of using Selenium, I analyzed the network traffic, identified the XHR requests that loaded the data in JSON, and simulated those requests directly with `requests-html`, a much faster approach. Once the data was obtained, I used Pandas to flatten, clean, and structure the result into a CSV.

🛠️ Tech Stack

📊 Impact & Results

The main result is a high-quality, citable dataset (DOI: 10.5281/zenodo.14043146) with 2,369 records, which can now be used by the community for economic analysis or price prediction models. The impact is the creation of a new public data asset.

🔗 Relevant Links

🎯 Context & Problem

Raw data on fatal traffic accidents (FARS) is rich but fragmented and full of inconsistencies. The problem was: how can we unify, clean, and enrich this data to create a single "source of truth" that allows data scientists to build reliable predictive models?

👤 My Role & Responsibilities

I took on the role of a Data Scientist focusing on the Data Understanding and Preparation phases of CRISP-DM. I was responsible for the entire ETL pipeline: joining tables, decoding variables, logical imputation of missing values, and creating new features.

⚙️ Technical Methodology

I transformed complex categorical variables into interpretable binary flags (`DRINKING`, `NIGHT_HOUR`) and created derived variables like vehicle age. I implemented a methodical cleaning of thousands of null values and special codes. Finally, I applied PCA to confirm the relevance of the new features.

🛠️ Tech Stack

📊 Impact & Results

The project delivered three clean, model-ready datasets. The impact is that this time-consuming preparation work is already done, allowing future efforts to focus directly on prediction and prevention to improve road safety.

🔗 Relevant Links

🎯 Context & Problem

The health-tech company Bellabeat needed to understand how consumers use tracking devices to identify market opportunities. The business problem was: What trends in Fitbit data can inspire new product features or marketing campaigns?

⚙️ Technical Methodology

The key technical highlight was the use of a hybrid pipeline: I processed most files in R with Tidyverse. However, for the large heart rate dataset, I uploaded it to Google BigQuery and used SQL to clean and aggregate it before re-importing it into R. I used ggplot2 for the visualizations.

🛠️ Tech Stack

📊 Impact & Results

The impact is measured by its final business recommendations. I proposed concrete actions such as a "sedentary alert" in the app or marketing campaigns focused on converting light to moderate activity, creating a clear bridge between data analysis and business strategy.

🔗 Relevant Links

🎯 Context & Problem

In unsupervised ML, choosing the right algorithm is crucial. The research question was: Can we demonstrate and quantify the superiority of a density-based algorithm like DBSCAN over K-Means on a real-world dataset that presents noise and irregular-shaped clusters?

⚙️ Technical Methodology

Nulls were imputed with a robust conditional mean. OPTICS was used to guide the parameter selection for DBSCAN. Quantitative validation was the core of the project, using the `fpc` library to calculate metrics like the Dunn Index and Silhouette Width. The results showed a drastic improvement in the Dunn Index (from 0.004 to 0.24), proving DBSCAN's superiority.

🛠️ Tech Stack

📊 Impact & Results

The impact is a clear, evidence-based demonstration of how to select the appropriate clustering model. The result is not just a set of clusters, but a deeper understanding of how and why these algorithms work, serving as a guide for other analysts.

🔗 Relevant Links

🎯 Context & Problem

A fictional company relied on static reports from its transactional database, making it impossible to answer complex business questions. The objective was to create an analytical data structure that would allow for "slicing and dicing" queries quickly and intuitively.

👤 My Role & Responsibilities

I took on the role of a Business Intelligence Analyst / Data Architect, responsible for designing the Data Warehouse schema (star model), creating the SQL scripts for the ETL, and finally, designing and configuring the Data Cube for analysis.

⚙️ Technical Methodology

The solution was based on a Data Warehouse with a fact table (`FactSales`) and several dimension tables (`DimProduct`, `DimCustomer`, `DimDate`, `DimLocation`), a fundamental structure for OLAP Cube performance that allows business users to explore data from any perspective without technical knowledge.

🛠️ Tech Stack

📊 Impact & Results

The project provides the company with a powerful Business Intelligence tool, democratizing data access and empowering managers to perform their own analyses, leading to faster, data-driven decision-making.

🔗 Relevant Links

🎯 Context & Problem

Obesity is a multifactorial problem. This project sought to answer: Can we, using only data on a person's habits, predict their weight category with a useful degree of accuracy and identify the most influential lifestyle factors?

⚙️ Technical Methodology

After a thorough Exploratory Data Analysis (EDA), the data was prepared and a range of classification models, including Logistic Regression, K-Nearest Neighbors (KNN), SVM, and Random Forest, were trained to determine the best performer for this public health problem.

🛠️ Tech Stack

📊 Impact & Results

The project resulted in a functional model that can serve as an educational or preliminary screening tool, identifying the most potent lifestyle predictors to help design more effective prevention campaigns.

🔗 Relevant Links

🎯 Context & Problem

For investors and regulators, early detection of bankruptcy risk is critical. The problem was: Can we build a model that serves as an early warning system and also provides clear, interpretable rules to justify its predictions?

⚙️ Technical Methodology

The choice of a Decision Tree was deliberate to ensure interpretability. This model allows for the visualization of the decision rules it uses (e.g., "if debt ratio > X and ROC < Y, then high risk"), which is crucial for trust in a financial setting. Performance was evaluated with a special focus on the Recall of the minority class (at-risk companies).

🛠️ Tech Stack

📊 Impact & Results

The project provides a decision support tool for financial risk management, enabling a shift from reactive analysis to proactive risk identification and timely corrective action.

🔗 Relevant Links

⚙️ Technical Methodology

The entire analysis was structured around a series of key questions, using Pandas for data manipulation and Matplotlib/Seaborn to create the visualizations that answered them. A significant part of the work involved preprocessing text data to reliably analyze genres.

🛠️ Tech Stack

📊 Impact & Results

The result is an analytical report that offers a data-driven snapshot of the television world. The impact is the demonstration of how EDA alone can generate value by converting data into knowledge and communicating complex insights simply and visually.

🔗 Relevant Links

⚙️ Technical Methodology

The project's strength was its creative feature engineering: I transformed `Cabin` into a binary `InCabin` flag, combined `SibSp` and `Parch` into `FamilySize`, and most importantly, extracted titles (`Mr`, `Mrs`, `Master`) from the `Name` column, revealing a powerful predictive signal. Five models (Logistic Regression, Decision Tree, Random Forest, k-NN, and SVM) were compared using the area under the ROC curve (AUC) as the primary metric.

🛠️ Tech Stack

📊 Impact & Results

The final SVM model achieved an AUC score of 0.82, demonstrating good predictive power. The real impact of the project is the demonstration of a robust and replicable classification workflow that I can apply to any other business problem, highlighting that great feature engineering often outweighs the most complex algorithm.

🔗 Relevant Links