Samuel López Santamaría

Graffiti is a problem in some regions of Germany, producing exorbitant costs for affected companies. The client, a regional train operator, had procedures in place to minimize recovery times, but none were preventive. I developed a machine learning model to predict upcoming graffiti on trains to prevent damage before it happens.

Using geographic and historical data, I led the development and testing of these models—XGBoost, LightGBM, and algorithms from scikit-learn. Also, I visualized the geographic distribution of graffiti hotspots with Kepler.gl and GeoPandas. The project further involved identifying key challenges with data quality and leading technical workshops to define countermeasures.

Swift turnarounds are key to an airline's operations. Wasted minutes on the ground can propagate quickly, generating a cascade of expensive, negative consequences for other flights. I participated in the development of a machine-learning-based approach to predict flight delays.

My team trained a model on 3D geographic data using PySpark and XGBoost, which outperformed the client's current delay prediction method. I also reduced computation time for one feature from days to minutes.

I led a team of six professionals, including data scientists, data engineers, and researchers. We used reinforcement learning to optimize the routing of self-driving buses in a village in Bavaria, Germany. I introduced an Agile workflow, coding guidelines, and other best coding practices. Further, I reviewed pull requests and guided development until a first-performing agent was found and trained with Stable Baselines3 (SB3). Also, I defined a development path to further increase the agent's performance.

Most wind turbine maintenance operations require the turbine to be shut down. Since the turbine cannot produce power, maintenance always implies lost profit. Ideally, maintenance would be carried out during low wind conditions to minimize this loss. However, scheduling is affected by a vast amount of constraints, and the task is not trivial.

I integrated into the team to develop a Python algorithm to compute optimized schedules. The underlying mixed-integer linear problem was solved with Gurobi and Python-MIP. Schedules were saved to a NoSQL database—Azure Cosmos DB. Data was gathered from a variety of internal APIs, while code quality was guaranteed by unit and integration tests. The algorithm was deployed as an Azure function.

I carried out user stories in all the steps described above, writing and reviewing code and coordinating development with the data engineering team.