Mathematics in Data Science / Machine Learning
I combine a strong mathematical foundation with applied machine learning, statistical evaluation, and careful communication. My work focuses on turning complex data into models and analyses that are technically sound, transparent, and useful in practice.
In this project, we implemented a neural-network-supported 3DVar data assimilation pipeline on the chaotic Lorenz ’63 system. We generated a synthetic twin experiment and computed innovations and 3DVar increments to establish a strong classical baseline. On top of that, we trained a compact MLP (ReLU, Adam, early stopping) to predict analysis increments directly from the background state and observations, and also evaluated a hybrid scheme that blends the learned increment with the 3DVar update. Evaluation included an 80/20 train–validation split, Monte-Carlo studies across multiple observation-noise levels, and a partial-observations setting. Across settings, the learned model consistently reduced mean L2/RMSE versus standard 3DVar. Overall, the approach demonstrated robust improvements in state estimation over classical 3DVar.
As part of a team of four, I worked on a project for Audi focused on predicting the time it takes to sell cars. We developed a deep learning model that accurately predicted the timespan, achieving a margin of error of just 9 days, significantly outperforming the baseline of approximately 20 days. To ensure the predictions were interpretable, we also utilized a Generalized Additive Model (GAM), which allowed us to gain valuable insights into the key factors influencing the sales times. The project placed a strong emphasis on handling high-dimensional data and performing comprehensive feature preprocessing to ensure optimal model performance. Our team presented our findings multiple times at Audi, demonstrating the model’s effectiveness and impact. The project was graded with a top score of 1.0.
In my second semester, I worked on a project focused on classifying news as fake or true of this Kaggle Fake and Real News Dataset, utilizing both classical machine learning classification methods and modern deep learning approaches. The project required extensive data preprocessing, feature engineering, and careful model evaluation to ensure accuracy and robustness. By applying a combination of traditional techniques and advanced neural networks, I was able to achieve highly reliable results. The final model achieved an accuracy of approximately 99%, demonstrating strong generalization across both classes. The project was graded with the highest possible score of 1.0.
