During my internship, I worked on a project aimed at characterizing the path of an event after a risk emerges. The focus was on critical materials, and the objective was to analyze textual data to provide deeper insights into potential outcomes based on previous detailed reports from the European Commission regarding why a risk takes a specific route. The project was divided into two key phases:1. Fine-tuning a RoBERTa model for sentence classification into four risk levels : Tasks included data scraping, chunking, and augmentation. I created data loaders and modules, and built the classifier by adding layers on top of the pre-trained RoBERTa model from Hugging Face. I configured hyperparameters, utilized early stopping, and trained the model with class-weighted loss to mitigate overfitting caused by imbalanced data.2. Developing a RAG (Retrieval-Augmented Generation) model to predict potential event trajectories: This phase involved indexing data from the European Commission and building a retrieval system (using all-MiniLM from Sentence Transformers for embedding). I augmented prompts with instructions and context, and for generation, employed a layer-wise inference technique to run LLaMA 3.1 70B on a 16GB RAM unit.

During my internship, I worked on data science projects focused on predicting currency liquidity and analyzing the factors affecting bank liquidity using advanced modeling techniques:1. Data Preprocessing: Cleaned and organized data for daily tracking of the cash balance sheet, ensuring high-quality input for analysis.2. Time Series Modeling: The goal was to predict currency liquidity and analyze the impact of various factors, such as interest rates, inflation, and exchange rates, on bank liquidity. I used multivariate time series models, including cointegration and causality analysis, followed by backtesting for validation.3. Classification Modeling: Developed a scoring model to classify and predict whether a bank treasury's balance would result in a debit or credit. This was achieved using logistic regression and XGBoost, with prediction calibration through the Platt method.