Five years of full-stack engineering. For the last two, I've been putting LLMs into the tax workflows of banks and multinationals: document reading, cost per call, and Slack hearing about drift before the client does.
Registering a folder of tax documents by hand took hours of analyst time. Now OCR plus OpenAI import 353 documents in 3 minutes and human review takes 10 more. Every call logs its own cost, and when a result drifts from the expected pattern, Slack hears about it before the client does.
I owned Taxly's analytics module from the first commit in June 2025 to production. Today it's 95 backend endpoints and five ready-made views that each enterprise client arranges its own way. Every chart is there because someone needs it to make a tax decision.
Before changing a line, I froze the behavior of all 84 endpoints with characterization tests and a semantic comparator. The migration ran in phases, and each phase shipped only when parity matched. No behavioral regression reached production.
Same period: a filter contract shared by 6 modules (FiltersV2) and a permissions rework across 340 files that surfaced a long-standing async race condition.
Owner of the analytics module from zero to production, integrations with dozens of government portals, and the 84-endpoint rearchitecture with proven parity. When something breaks, I fix the cause, not the symptom.
AI-based website and app generators, WhatsApp automation, and the migration of the codebase to Vue 3 + Nuxt with broad test coverage.
“Consistent, above-average delivery and results. Any team is lucky to have him on board.”
“His ease getting along with others and his brilliance as a programmer set him apart on any team.”
A platform for US nonprofits: invoices, proposals, donations and sweepstakes, with Stripe integration and a UI/UX redesign.
Logistics and e-commerce APIs, ERP dashboards, and the habits of keeping code alive in production.
Ask about Brazilian tax law or live company data, and every claim points at the exact span or API response behind it. A second model audits each claim, deterministic gates audit the auditor, and a question the evidence cannot support gets refused. Built to practice RAG, tool calling and LLM evals.
Pull fields out of documents without reviewing every one by hand: a queue, retries guided by the model's previous error, and human review only where confidence drops. Built to practice dead-letter queues, versioned evals and LLM guardrails.
Real-time 1v1 Pokémon battle simulator with a server-authoritative engine that validates every move. Built to practice WebSockets and a real deploy: AWS ECS Fargate through GitHub Actions.
On top, what I use every week. Below, what I've already shipped to production.
SAP office · São Leopoldo, RS · Jun 2026
Unisinos · Porto Alegre, RS · Dec 2025
University of Caxias do Sul (UCS)
IFRS · Bento Gonçalves Campus