Andrea Beltran Lizarazo: Colombian Extraditions to the US

The media often spotlight high-profile extraditions of wealthy and violent drug lords—mostly from Latin America, especially Mexico and occasionally Colombia—while overlooking a more common but less sensational reality: the majority of drug-related extraditions to the U.S. likely involve low-level offenders. Andrea’s dissertation uses Colombia as a case study to document and explain this broader trend. In past terms, Spark teams have successfully extracted data from more than 2,500 extradition cases spanning two decades, relying on court files and executive documents from official Colombian sources. This term, we will build on that work with two main missions. Each presents technical challenges that will allow the team to practice and expand core data science skills, while also contributing directly to Andrea’s research: 1. Corpus: PDFs to dataset: We have a collection of roughly 600 PDF documents, each about 40 pages long, which include short official communications of approximately one page each. The task is to extract these communications and organize them into a structured spreadsheet. The team will apply techniques in text mining, PDF parsing, and data structuring, and Andrea will provide textual patterns to define the start and end points for extraction. • Learning opportunity: working with large-scale text extraction from complex documents, building automated workflows for data cleaning and structuring. 2. Entity extraction with NLP/LLM tools: We also have an existing dataset that contains references to attorneys. The task is to process this dataset using NLP techniques, including an LLM agent, to extract entities such as attorney names and roles. The team will need to address challenges such as typos, inconsistent formatting, and linguistic variability. A preliminary attempt achieved an accuracy rate of about 66–67%. With improved inputs and refined methods, the mission is to raise that level of accuracy. • Learning opportunity: applying natural language processing to messy real-world data, experimenting with LLM pipelines, and refining extraction accuracy.