Leon – LLM, Python, AWS

A full-scale asynchronous platform developed to automate the classification and extraction of critical corporate documents. Unlike standard keyword-matching tools, this system uses AI to evaluate the purpose and function of documents, effectively handling regional terminology, abbreviations, and synonyms across Brazil, Mexico, and international markets. It serves as a central hub for compliance, replacing manual, labor-intensive verification processes that previously drained valuable work hours.

• Architected a performant async platform capable of classifying documents based on core function rather than simple text patterns.
• Developed a rendered frontend for simplified and direct product team usage of the solution.
• Implemented regional classification logic for 6 specific document types, including Brazilian, Mexican, and some global documents.
• Designed the multi-agent AI system that interprets complex legal terminology to ensure accurate categorization across diverse LATAM jurisdictions.
• Reduced manual curation costs and improved evaluation speed for massive document volumes, directly impacting SLA reliability.
• Deployed automated CI/CD pipelines using GitHub Actions, ensuring safer tagging and versioning for high-frequency updates.

A service curating both public and private sources for risk and false positive entity-related information. The application addresses companies and people from all around LATAM and international countries to check sources, classifying their compliance risk and generating a scoring that determines if the entity is indeed related to the entity identified in the article, thus speeding the evaluation of massive volume of articles for thousands of customers, reducing the cost of curating, checking and classifying those content manually and easing up the access of multiple combined of sources.

• Designed the back-end, front-end section, and the AI systems used in the current product;
• Created the integration of multiple data sources by adding a common interface that supports multiple article sources;
• Worked on a new Scraper to extract public articles (extracting more than 10k articles daily);
• Created a database that holds more than 100k entity-articles evaluated by the AI system;
• Deployed solution using CI/CD in git actions and an automated versioning system for safer tagging;
• Impacted over 4 reports and 1 new API pipeline under testing for data|AI as a service. (LATAM and International)

A specialized data analytics and profiling platform created to identify the distinguishing factors between high-performance and low-performance student categories. The platform extracts raw educational data, profiles student behavior, and applies unsupervised learning to discover hidden clusters within the student body. This enabled the company to move beyond surface-level metrics and understand the underlying drivers of academic success within its institution.

• Created a sliced student profiling system that categorized students into 9 quadrants (from lowest to highest performance) based on class presence and activity execution.
• Developed an unsupervised model using K-Means clustering to extract distinct clusters of student behaviors automatically.
• Built interactive visualization tools with Streamlit, providing stakeholders with clear exploratory and explanatory data reports.
• Deployed the solution in a secure VM environment, providing a friendly and restricted interface for authorized Sagres users.

An intelligent NLP solution designed to process real-time financial news streams using natural language processing (NLP). The system focuses on named entity recognition (NER) and multi-label classification to tag and categorize market-moving information automatically. By leveraging TC’s private datasets, the platform surfaces relevant news with minimal latency, providing a significant competitive advantage in financial data analysis and enabling our clients to customize news retrieval.

• Developed an NLP architecture for multi-categorization and multi-labeling of high-frequency financial news.
• Built and optimized a multi-label classification system to provide intelligent data-driven insights to users.
• Curated training datasets using Argilla, leveraging private data to improve the accuracy of financial entity recognition.
• Collaborated with the ML team to build end-to-end intelligent solutions that transformed raw text into actionable financial intelligence.