I decide where AI belongs in an operational pipeline, and where it does not. What becomes an agent, what stays deterministic code, and what should not be built at all.
Currently at Banco Inter's Global Operations, where I design the system prompts, tools and MCP integrations behind agents running in production, plus the Python automations around them. Nine of those shipped in seven months.
Working inside a bank means every agent ships under audit, access and governance constraints, so I care as much about evaluation and guardrails as about the prompt.
Alongside that I do undergraduate research at CEFET-MG on hybrid LLM models for market prediction and financial risk, where the point is benchmarking the combined method against each approach on its own rather than reporting one number in isolation.
Previously at Akyou, where I built an end-to-end document intelligence pipeline using the OpenAI API.
·Took a document intelligence product from scratch to a working prototype, owning it end to end
·Built the Python + OpenAI backend and LLM pipelines for unstructured PDF processing, extracting and restructuring complex documents into accessible, structured output
·Built the TypeScript frontend from the ground up and deployed the backend on AWS Elastic Beanstalk behind a load balancer
2024 to 2030 (expected) · Currently in the 5th semester
Projects
Vale Desenvolver 2026: Mining Fleet Telemetry
Deep dive
End-to-end predictive pipeline to anticipate critical failures (Don't Go events) on 37M+ telemetry records from an iron ore mining fleet. Submitted to Vale's Programa Desenvolver 2026, an open challenge.