About Guilherme Inácio

Who I am

I am Guilherme Inácio, a Computer Engineer trained at the Federal University of Santa Catarina (UFSC) in Brazil, with two semesters of Informatics at the Technische Universität München. I live and work in Florianópolis, Santa Catarina, and I have been building software with data and machine learning at its centre for roughly six years.

Today I work as an AI Engineer at Abstrato, an enterprise AI consultancy, designing agentic architectures and the infrastructure that keeps them reliable in production: tool servers, orchestration, evaluation and observability.

What I work on

My focus is the plumbing that makes large language models genuinely useful: giving them well-typed tools, trustworthy context and guardrails that hold under real traffic. In practice that means Model Context Protocol servers, LangGraph and Claude Agent SDK orchestration, retrieval pipelines over PostgreSQL and pgvector, and the data engineering underneath.

• AI agents and orchestration: LangChain, LangGraph, Pydantic AI, Claude Agent SDK.
• Model Context Protocol: FastMCP and TypeScript servers for Google Workspace, Microsoft 365 and more.
• Data and analytics: Python, SQL, PostgreSQL, BigQuery, dbt, pandas, Streamlit dashboards.
• Infrastructure: FastAPI, Docker, Google Cloud Run, CI and deployment automation.

Career so far

At Franq I built an automated validation and document-analysis system that removed most of a manual verification workload, an evaluation pipeline that compared model output against gold-standard datasets with drift detection, and LLM observability with LangFuse. Before that, as a Data Analyst, I delivered an incentive-campaign platform that centralised performance monitoring across sales and customer success teams.

At Meton Soluções I ran business intelligence for around fifty dealership groups, building commission dashboards and improving query performance by more than three hundred percent through indexing and normalisation. In 2019 I interned at the Czech Technical University in Prague, applying statistics and machine learning to EEG data in support of epilepsy surgery research, and continued the collaboration remotely as my undergraduate final project.

Open source

I publish most of my MCP work openly on GitHub: servers that expose Google Workspace and Microsoft 365 as typed tools, a Streamlit playground for connecting any model to any MCP server, an enhanced memory server with graph queries, and healthcare analytics agents over Brazilian public health data. This website itself is open about its agent interface; the tools, policies and limits are all documented at /developers.