I build AI and data systems that check their own numbers.
Data and Machine Learning Engineer (GenAI and LLMs) in Rio de Janeiro. I take LLM products to production and build the AWS and Google Cloud data platforms behind them. Certified on Google Cloud and Oracle Cloud (OCI).
Open to remote roles, full-time or contract, with US or European teams.
The audit behind dental-analytics-medallion: services per patient fell from 10.41 to 6.34 once duplicate rows and suppressed counts were handled. Point at a model to trace what it reads from.
Open-source projects
Each one is tested and type-checked (mypy strict) in CI. The result beside it is something I measured.
-
docfunnelpypdfium2, RapidOCR, Pydantic, FastAPI, MCP
Turns PDFs and scans into data an LLM can be trusted with: fast text extraction, OCR only on the pages that need it, schema-validated extraction with retries and model fallback, and evals that track accuracy, invented fields, cost and latency. Usable from a CLI, an API or as an MCP tool.
Sends a page to OCR only when needed, right 99.3% of the time, and reads light scans with a 1.6% character error rate.
-
lgpd-face-redactorOpenCV (YuNet, SFace), FastAPI, Docker
Blurs every face in a video except the people you choose, for LGPD and GDPR access requests. Upload the footage, tick who stays visible, then download the redacted copy and an audit record. Runs on your own machine.
99.38% verification accuracy on LFW, and a policy that blurs whenever a match is in doubt.
-
dental-analytics-medalliondbt, DuckDB, OpenAI structured outputs, Streamlit
A bronze, silver and gold pipeline over California's public dental data, with a dashboard and an LLM brief whose every figure is checked against the warehouse before anyone reads it.
The audit found services counted twice:
71.2 millionbecame 42.6 million, and 505 duplicate rows came out. -
marketing-report-generatorpandas, statsmodels, SciPy
Funnel and channel analysis of messy CRM exports, reporting intervals and corrected significance tests instead of bare percentages.
Checked against known true rates on 200 simulated exports per size: it flags the better channel in 89% of them at 5,000 prospects, and in the wrong direction at most 1%.
-
uerj-marica-transit-networkNetworkX, MapLibre
A graph model of the public bus network of Maricá, RJ: centrality, shortest routes, reach by number of transfers and stop spacing, with an interactive explorer in Portuguese and English.
614 stops and 46 lines. Removed a fake hub created when 33 stops were geocoded to one point.
-
uerj-optimization-methodsNumPy, SciPy, Pyodide
Nelder–Mead written from its definition, and a constrained tank-design optimizer using Newton and quasi-Newton methods, both runnable in a browser lab.
Matches SciPy's evaluation count on every benchmark. The original solver reported an infeasible tank as the optimum; now it doesn't.
More coursework, rewritten as tested packages: finite-difference and TVD advection schemes, search, constraint satisfaction and an MLP from scratch, and least squares with inference and diagnostics.
Work
-
to present
Data and Machine Learning Engineer
A government-sector organization
- Migrated the transformation layer from Spark on EMR clusters to SQLMesh on Athena, orchestrated by Airflow, with incremental models, column-level lineage and tests on every model: about 40% faster at about 65% lower cost.
- An LLM pipeline that reads submitted documents, extracts the required fields with schema-constrained outputs and pre-fills a form, with a human review before anything is saved.
-
to
Machine Learning Engineer, GenAI and LLMs
IPNET by Vivo
- A WhatsApp customer-service agent, rolled out to about 1,000 businesses, that answers with business actions and a knowledge base, keeps each customer's context in the CRM and hands off to humans through email tickets or Google Cloud CCAI.
- Accelero, a Google Chat assistant that turns the last 24 hours of conversations into each person's action items every morning. Messages are processed in memory only.
- Infrastructure as code with Pulumi that provisions governed Google Cloud projects end to end.
-
to
Data Engineer
IPNET by Vivo
- Migrations of tables over 1 TB for a large Latin American retailer, with Cloud Composer and Dataflow.
- A three-layer data lake for a government client, with data quality checks.
- Waze traffic pipelines feeding urban monitoring dashboards in Looker.
-
to
Director of Computing Projects
Serra Jr. Engenharia
- Scoped and budgeted client software projects, led the teams and wrote code alongside them.
Toolbox
- Languages
- Python, SQL, TypeScript and Node.js, Go
- AI and ML
- LLM orchestration, agents and MCP servers, RAG and embeddings, structured extraction, evaluation, Vertex AI, CCAI, computer vision
- Data
- SQLMesh and dbt, Airflow and Cloud Composer, Beam and Dataflow, BigQuery, Athena, Snowflake, DuckDB
- Backend and infra
- FastAPI, Docker, Kubernetes, Terraform, Pulumi, Kafka, GitHub Actions, Azure DevOps
- Cloud
- AWS (Athena, EMR, ECS, RDS, EC2), Google Cloud, Oracle Cloud Infrastructure (OCI)
Certifications and education
- Google Cloud
- Associate Data Practitioner, Cloud Digital Leader
- Oracle Cloud (OCI)
- Professional in Generative AI, Data Science, DevOps and Developer; Associate in Foundations, AI Foundations and Data Platform
- Degree
- BSc in Computer Engineering, UERJ (2020 to 2025), where I was a Calculus I teaching assistant
- Startup
- Co-founded Opinei and took it from idea to investor presentations (TIM Brasil Academic Working Capital, 2022)
- Languages
- Portuguese (native), English (C1), Spanish (advanced)
Hiring for a data or ML role?
I'm open to remote roles, full-time or contract, with US or European teams. The quickest way to reach me is LinkedIn, and the code for everything above is on GitHub.