Matheus Rosa

Experience

Maggu — Remote

AI Engineer · Apr. 2026 – Aug. 2026

• LLM evaluation
  Improved medicine recommendation accuracy from 62% to 94% by replacing a naive LLM judge with two adversarial agents backed by a human-made dataset in MLflow on Databricks.

• Observability and latency
  Tracked LLM costs, latency, accuracy, and output routing in Databricks MLflow. Reworked caching and RAG to cut time to first token from 1.5 to 0.7 seconds.

• Data infrastructure and performance
  Provisioned reproducible Databricks pipelines with Terraform. Reduced PostgreSQL upload time from 2 hours to 15 minutes with batching and eight upload threads; cut Databricks compute spend from $3.21 to $0.80 per hourly DBU by optimizing PySpark partitioning, joins, and small files.

Stack: Databricks, PySpark, Delta Lake, MLflow, Terraform, Agno, GitHub Actions.

Capgemini — Recife, Brazil (Hybrid)

AI Engineer · Jan. 2026 – Apr. 2026

• Agent evaluation
  Built an evaluation suite checking tool choice, arguments, and call order in agent workflows, catching four regressions before they reached the client.

• Team enablement and client delivery
  Ran workshops on agent design and LLM-as-a-judge for a new team and two Vivo managers. Handled technical client communication and mentored two data managers on agent evaluation.

Stack: Python, LLM evaluation, agent frameworks, tool-calling validation, RAG, Agile/Scrum.

Triggo.ai — Remote

AI Engineer · Dec. 2025 – Jan. 2026

• LLM output control
  Used self-consistency and tree-of-thoughts prompting to keep tone consistent. Tuned Anthropic and OpenAI API calls for JSON-formatted, length-constrained outputs.

Stack: Python, Bruno, DocumentDB.

Oncase — Recife, Brazil (Hybrid)

Data Engineer · Aug. 2025 – Nov. 2025

• Computer vision inference
  Maintained a Flask image-processing API with S3 uploads for a computer vision inference pipeline.

• Orchestration
  Maintained Airflow DAGs using a Gitflow branch pattern to keep broken DAGs out of production. Hosted Airflow on EC2 for continuous inference and training jobs.

Stack: Python, Flask, PaddleOCR, Airflow, AWS, Terraform.

Acaso — Recife, Brazil

Machine Learning Engineer · Jan. 2025 – Jun. 2025

• Document ingestion and RAG
  Automated document ingestion in Airflow, giving the team back roughly 20% of its week. Built a RAG pipeline with Docling for PDF parsing, LangChain for orchestration, and PostgreSQL with pgvector for storage.

• Cloud ingestion and prototypes
  Cut manual data preparation by 60% by automating S3 ingestion with Python and Boto3. Shipped generative AI proofs of concept in Streamlit and FastAPI for client validation.

Stack: Python, LangChain, PostgreSQL/pgvector, Docling, Apache Airflow, AWS S3, Boto3, FastAPI, Streamlit.

Di2win — Recife, Brazil

Machine Learning Engineer · Apr. 2024 – Jan. 2025

• Energy forecasting
  Cut energy consumption by 20% with LSTM and regularized-regression forecasts tuned in Optuna, outperforming a SARIMA baseline.

• ML delivery
  Replaced manual retraining with event-driven Airflow pipelines triggered by data refresh and drift. Added GitHub Actions and Pytest coverage for model behavior and input data; isolated training and retraining from live projects in containers.

Stack: Python, PyTorch, Optuna, Apache Airflow, Docker, GitHub Actions, Pytest, FastAPI.

Valorian — Recife, Brazil (Remote)

Data Scientist · Apr. 2023 – Apr. 2024

• Cloud and data automation
  Reduced hosting costs by 25% through containerization and reserved instances. Automated end-to-end ingestion and transformation, eliminating 30 hours of manual work each month.

• Production optimization
  Improved wheat flour production efficiency by 12% with XGBoost models tuned in Optuna.

Stack: Python, XGBoost, Optuna, Docker, AWS EC2/S3, Dash/Plotly.