matheus sampaio

Improved medicine recommendation accuracy from 62% to 94% by reworking the LLM judge into two adversarial agents backed by a human-made dataset on MLflow hosted at Databricks. Tracked LLM operational costs, latency, and accuracy, and reworked cache and RAG strategies to cut time-to-first-token from 1.5 to 0.7 seconds.

Wrote evaluation suites for agent workflows, catching 4 regressions before client delivery. Ran workshops on agent design and LLM-as-a-judge for a newly formed team and two managers from Vivo, and handled client technical communication in an Agile delivery team.

Engineered prompts to keep LLM tone consistent using self-consistency and tree-of-thoughts techniques. Tuned Anthropic and OpenAI API calls to enforce JSON-formatted, length-constrained outputs.

Maintained a Flask API for image processing, storing uploads on S3 for a computer vision inference pipeline. Maintained Airflow DAGs using a Gitflow branch pattern and hosted Airflow on EC2 to run continuous inference and training jobs.

Automated document ingestion in Airflow, giving the team back roughly 20% of its week. Built a RAG pipeline using Docling for PDF parsing, LangChain for orchestration, and PostgreSQL with pgvector for storage, and cut manual data prep by 60% by moving S3 ingestion into Python and Boto3.

Cut energy consumption 20% with LSTM and regularized regression forecasts tuned in Optuna. Replaced manual retraining with event-driven Airflow pipelines and set up GitHub Actions with Pytest coverage on model behavior and input data.

Reduced hosting costs 25% by containerizing workloads and moving to reserved instances. Automated ingestion and transformation end to end, saving 30 hours of manual work per month, and improved wheat flour production efficiency 12% with XGBoost models tuned in Optuna.

Pursuing a Bachelor of Science in Computer Science at the Center of Informatics, Federal University of Pernambuco, with expected completion in July 2026.