I build and govern the data platform that makes your AI work in production.

The foundation the rest of the portfolio is built on. Public Brazilian macro and market data lands as Delta on MinIO, is conformed by Spark into a star schema and served as DuckDB marts — every lane generated from one source registry. Adding a series is a YAML entry, not new code.

• 51 registered series across 10 source handlers (BACEN, IPEA, IBGE, Tesouro, ANBIMA, B3, Yahoo Finance) — medallion bronze → silver → gold on Delta Lake over MinIO, ending in 8 BI-ready marts.
• Airflow 3 Asset scheduling derived from the registry at parse time — 13 DAGs and 51 bronze assets, one task per series, so a failing API withholds exactly one asset and raises exactly one attributable alert.

• Spark
• Structured Streaming
• Delta Lake
• dbt
• Airflow 3
• MinIO
• Polars
• DuckDB
• Pandera
• OpenLineage
• OpenMetadata
• Kubernetes

The Kubernetes/GitOps layer that deploys this stack lives in a separate repository that holds cluster credentials. It stays private permanently and is not linked from here.

RAG over 30 public minutes of the Copom, Brazil's monetary policy committee, built around an evaluation harness rather than a demo. Thirty documents each contain a near-identical decision paragraph, so dense retrieval confidently returned the right paragraph from the wrong meeting. Seven measured retrieval arms later, that defect is gone — and the component that fixed it was not the one anybody would have bet on.

• The defect, measured: the dense baseline scored MRR 0.191 and got the right meeting at rank 1 on 4 of the 41 questions that name their meeting outright. It was not confused about the topic — it was answering a question nobody asked.
• The fix, measured on the same question set: MRR 0.191 → 0.741, hit@5 0.959, rank-1 meeting correctness 0.098 → 1.000. It came from ~120 lines of regex that read the date out of the question and compile it into a Qdrant payload filter, at 41/41 hint precision and negative latency cost.

• Python
• Qdrant
• bge-m3
• BM25
• cross-encoder
• Ollama
• Presidio
• Langfuse
• FastAPI
• DuckDB

Every number above comes from a committed JSON report in the repo. The gold set behind them is 56 draft pairs, not human-validated — so read the movement between arms, not the absolute level.