Federico – LLM, NLP, Python

LLM-powered chatbot application to leverage client’s internal database. The application showcases a first agent to translate natural language prompts into the client’s specific query format through prompt engineering, and a second one to summarize the returned data for the user.

• Worked on POCs of generative AI solutions for unstructured document analysis, leveraging OpenAI’s pre-trained LLMs and Azure’s cognitive services to determine valuable information regarding tax compliance;
• Included OCR of handwritten legislative documents and a chatbot assistant for tax analysts concerning exemptions and tax rates, utilizing an RAG architecture and an Azure Cosmos DB vector database built from the scraping of official sources;
• Designed and coded a framework to enable the usage of an API through a natural language interface;
• Developed a Python framework to facilitate the operation of the client’s API, allowing retrieval of information from a 270-million-organization database via a chatbot interface.

Chatbot assistant for tax analysts concerning exemptions and tax rates, powered by an RAG architecture and an Azure Cosmos DB vector database built from the scraping of official sources.

• Was in charge of building a RAG-powered chatbot assistant to provide accurate responses to users on highly technical topics.

Fraud detection algorithm for transactions on online money transfers for a big oil company.

• Designed and coded entire architecture for automated machine learning pipelines;
• Conducted feature engineering, model training, evaluation, and deployment;
• Developed machine learning pipelines that automatically adjusted to users' data, including diagnosing uploaded datasets, performing deep feature engineering, evaluating instantiated models, inferring on future data, and translating results into actionable insights;
• Demonstrated a strong focus on code optimization, profiling, and deployment of models into production environments.

Fraud detection algorithm for online transactions for a big oil company.

• Conducted data ingestion and implemented ETL processes to preprocess raw data for analysis;
• Engineered features to extract relevant information and enhance predictive model performance;
• Developed machine learning models to detect fraudulent transactions using highly imbalanced historical data;
• Evaluated model effectiveness and fine-tuned parameters to improve performance;
• Managed the development process under heavy inference time constraints to ensure timely predictions.

Clustering and recommendation system for a medium-sized food distribution company.

• Conducted feature engineering to enhance data quality and extract meaningful insights;
• Developed machine learning models for predictive analytics and recommendation systems;
• Evaluated model performance and translated results into actionable marketing strategies;
• Designed and implemented two sequential algorithms: one for clustering the company's customers and another for creating a personalized recommendation system for the product catalog;
• Utilized algorithms to deepen understanding of client consumption patterns and identify opportunities for sales growth.

Sports results’ prediction.

• Collected data via web scraping;
• Conducted ETL (Extract, Transform, Load) processes;
• Performed descriptive statistical analysis;
• Developed machine learning models to predict results of sports competitions, primarily horse racing;
• Engaged in feature engineering, experimentation, and research on a daily basis.

Tax Returns and accounting advice.

• Calculated tax returns for individuals and provided tax advice to employees of international companies who were temporarily or permanently relocated internationally;
• Developed online statistical reports on historical and future employee leave permits to track the department's available resources;
• Was in charge of generating data on the department's personnel and constructing Power BI reports with descriptive statistics and visualizations.