Joseph Martínez

• Created a custom conversational LLM (akin to ChatGPT) fine-tuning Llama2 with a dataset of 330 news.
• Trained Machine Learning models (KNN, RFC) and fine-tuned NLP models (BERT) to classify frustration types.
• Designed and maintained detailed documentation of models, algorithms, and reports on GitHub repositories.
• Presented poster at a research conference in Pittsburgh, and attended others in Arlington, VA, and Oxford, OH.

• Geolocated, filtered, and estimated the tone of 7.3 million Tweets through fine-tuning fifteen NLP models (spaCy, BERT, RoBERTa) and Statistical Analysis. Models were highly reliable, with a mean accuracy of 72%.
• Extracted, cleaned, and pre-processed 15.5 million Tweets in multiple languages with Twitter’s API and Python.
• Created dashboards with Tableau and Google Data Studio to represent migration data.
• Collaborated with teams from Norway, Colombia, Greece, and the US on a $1.4M Minerva research initiative.

• Trained a Natural Language Understanding (NLU) model with IBM Watson to semi-automate the identification of actors, factors, and relationships from news articles, achieving a final accuracy of 70%.
• Gathered, filtered, and cleaned geo-tagged 4,680 Tweets on migration using Twitter's API and Python.
• Conducted a lexicon-based Sentiment Analysis of Tweets to spot xenophobic trends.
• Determined key actors of a Twitter conversation about migration through Social Network Analysis with Gephi.