Experience

Oncase — Recife, Brasil

Data Engineer · Sep. 2025 – Dec. 2025

• Document data extraction
  Enabled more reliable and accurate document understanding for the client’s product by improving how text is extracted from real-world documents, allowing the solution to adapt to data variability and continuously improve over time.
  Implemented using PaddlePaddleOCR.

• Stack used
  Delivered the above solutions using a scalable orchestration and processing stack.
  Airflow orchestrating AWS ECS tasks on a daily schedule, with PaddlePaddleOCR for document text extraction.

Acaso — Recife, Brazil

Data Scientist · Jan. 2025 – Aug. 2025

• Reliable data ingestion pipelines
  Reduced time spent on data ingestion issues and increased focus on model development by building stable, automated database pipelines for structured data.
  Implemented PostgreSQL ingestion handlers in Python using Psycopg2.

• Retrieval-augmented generation systems
  Enabled accurate and context-aware information retrieval from large document collections, improving the quality of LLM responses for internal and client-facing use cases.
  Built RAG pipelines with LangChain and PostgreSQL VectorDB, generating embeddings from PDFs processed via Docling.

• Rapid LLM prototyping
  Accelerated client feedback cycles by delivering interactive LLM-based prototypes early, allowing validation of use cases while preserving data privacy and database consistency.
  Developed LLM-based POCs using Streamlit and FastAPI.

• Automated document ingestion
  Ensured timely and reliable data availability by automating document ingestion workflows, enabling teams to start each day with fresh, processed data ready for use.
  Scheduled ingestion pipelines using Apache Airflow. Saving about 6 hours weekly from working hours.

• Cloud storage integration
  Reduced manual data handling and operational overhead by enabling seamless file ingestion and management, saving approximately 60% of the time previously spent on manual data preparation.
  Built AWS S3 integrations using Python and Boto3, supporting CLI and Streamlit-based execution.

Di2win — Recife, Brazil

Data Scientist · Jun. 2024 – Jan. 2025

• Production ML data pipelines
  Enabled reliable machine learning in production by converting raw data into high-quality, versioned features that support training, validation, and inference across the full ML lifecycle.
  Built end-to-end ingestion, cleaning, and feature engineering workflows.

• Model deployment and optimization
  Improved operational efficiency by deploying and continuously optimizing machine learning models tied to business KPIs, enabling strategic reordering of operations and reducing energy consumption by approximately 20%.
  Deployed regularized regression and LSTM time-series models, optimized with Optuna.

• Automated training and retraining pipelines
  Accelerated experimentation and reduced manual overhead by automating training, retraining, and data refresh workflows, allowing models to adapt to new data or events without manual intervention.
  Implemented event-driven and scheduled pipelines using Apache Airflow.

• ML reliability and production readiness
  Increased production stability and model trust by enforcing automated testing, validation, and deployment checks, resulting in fewer production incidents and safer releases.
  Implemented CI/CD pipelines with GitHub Actions and model/data tests using Pytest.

• ML platform and delivery strategy
  Enabled fast and safe model delivery while protecting Di2win’s intellectual property by designing an ML platform that supports frequent retraining, continuous deployment, and parallel experimentation without disrupting active projects.
  Designed a scalable, isolated ML stack for rapid iteration.

Valorian — Recife, Brazil

Data Scientist · Apr. 2023 – Apr. 2024

• Flexible data management
  Supported rapid product iteration by designing a data management approach that allows continuous processing and frequent changes without disrupting the system, giving the client freedom to refine and validate their MVP. Saving 30% of its budget adapting to the new format asked.
  Backed by a flexible data storage design.

• Automated data processing workflows
  Reduced operational effort and improved reliability by automating end-to-end data ingestion and transformation, enabling the product to scale and evolve smoothly while saving approximately 30 hours of manual work per month.
  Automated using scheduled workflows and containerized tasks and reserved instances on AWS, saving about 20% to 25% per month on budget.

• Interactive dashboards
  Developed dashboards using Dash Plotly for clients so they understand the storytelling behind the data they’ve provided.

• Machine learning models optimized with Optuna
  Optimized XGBoost models, improving cost efficiency in wheat flour production by 12% by identifying the production orders that yield the greatest benefit.

Highlights

• Impact-driven work
• Strong collaboration
• Continuous learning