Nelson Joppi

I’m an AI Engineer focused on scalable LLM platforms, agentic workflows, and production-ready data pipelines. I’m the core engineer behind VigIA, an award-recognized auditing platform that helps flag irregularities across R$ billions in procurement documents, and I’ve contributed to a revenue-impacting sales productivity copilot at Stone.

At the Court of Accounts of Santa Catarina, I led the redesign from RAG-only Q&A to a multi-task LLM platform for non-technical auditors. I built a generic LLM API plus a multi-step workflow executor (Q&A, summarization, classification, extraction), integrated GCP/Azure models, and productionized orchestration with Airflow + Spark/Hadoop—owning maintenance, code reviews, and the service/pipeline evolution.

At Stone, I conducted field discovery for Manezinho (a WhatsApp assistant for sales agents), shadowing agents, running 1:1s, and translating needs into workflows/intents and a prioritized backlog. I proposed hybrid intent routing (rules/regex fast-path + LLM) to improve quality while reducing latency/cost, and I designed statistically rigorous experimentation for revenue-impacting agent tools—defining KPIs, randomization strategy, variance estimates from historical funnels, and controls for time trends.

Earlier, at FEESC I grew from Research Assistant to ML Solutions Engineer, maintaining an ML plugin used by Petrobras geologists to train and apply lithology models end-to-end on well-log data (reaching ~80% accuracy in representative evaluations). I also built ML usability for domain experts through onboarding, UX/workflow iteration, and shipped grid search with safe defaults and guardrails—before that, I delivered React apps with Node.js APIs at Voltbras and developed a genetic algorithm for road network design during undergraduate research.