I'm a Data Scientist based in Rio de Janeiro, Brazil 🇧🇷. Currently, I work at FGV IBRE, focusing on Generative AI, Machine Learning, and Data Engineering within the AWS ecosystem. My work involves building advanced NLP pipelines, developing classification models, and optimizing cloud data processes using Snowflake and Streamlit.

I hold a B.Sc. in Computer Science and have a strong background in automating business processes and BI from my previous experiences at Bagaggio and Enel.

Fun fact: Before diving into data, I was a professional e-sports player, a journey that sharpened my resilience, strategic thinking, and ability to perform under high pressure 🎮.

Languages & Frameworks

Cloud & Data Engineering

Vector Stores & AI Infra

Streamlit platform that automates press release generation for 15+ FGV IBRE economic indicators. Ingests Excel data via pandas/openpyxl, assembles few-shot prompts from historical bulletins, and calls Azure OpenAI (GPT-4o) to generate narratives. Final output is rendered as a formatted Word (.docx) via docxtpl (Jinja2) and logs downloads to Snowflake. Optionally enriches context via a RAG API before generation.

End-to-end semantic search and RAG system for Portuguese-language journalism and economic articles. The pipeline fetches from MongoDB/AWS DocumentDB, cleans noisy text with spaCy (pt_core_news_sm) + NLTK, generates dense embeddings via sentence-transformers, and stores them in Qdrant with SHA-256 deduplication and checkpoint/resume support. A LangGraph-orchestrated framework then enriches economic analysis text by querying Qdrant across multiple collections, scoring chunk relevance with an LLM, and integrating context via Azure OpenAI (GPT-4o) — with configurable quality thresholds and YAML prompt overrides. Built for AWS SageMaker deployment.

🌐 Live app — End-to-end RAG system for Q&A over COPOM (Brazilian Monetary Policy Committee) documents. The pipeline downloads meeting minutes and communications from the BCB Open Data API, parses PDFs with pdfplumber, chunks with LangChain + tiktoken, embeds with Google Gemini, and stores 1536-dim vectors in PostgreSQL + pgvector (HNSW index). A FastAPI service retrieves chunks via cosine search, optionally reranks with LLM, and generates cited answers with Gemini. A Streamlit app delivers the Q&A interface, document filters, and an admin panel with live ingestion logs.

Monthly economic activity monitor for all 5,570 Brazilian municipalities derived from VIIRS/DNB nighttime satellite imagery (NASA/NOAA). Batch-extracts radiance data via Google Earth Engine, applies Winsorization (p99), IHS transform, and temporal Z-score normalization per municipality, then loads into Snowflake. An interactive Streamlit + PyDeck WebGL dashboard renders choropleth maps, municipal time series, and regional rankings.

Automated stock analysis system that scrapes fundamental indicators from Investidor10 via BeautifulSoup4, stores them in PostgreSQL, and trains scikit-learn Random Forest models for both price forecasting (regressor) and performance classification. A daily schedule-based orchestrator runs scraping, multi-day predictions, and recommendations automatically. Results are explored through a Dash/Plotly dashboard with Bootstrap theming, backed by Docker Compose.

End-to-end music genre classification pipeline: ingests tracks from YouTube via yt-dlp, stores metadata in MongoDB, and extracts 369 audio features per track using librosa (MFCCs, Chroma, Spectral Contrast, ZCR, RMS, Tempogram). Trains XGBoost and Random Forest classifiers, and explores clusters via PCA → K-Means → t-SNE. A Streamlit app accepts YouTube links or file uploads and returns real-time genre predictions with waveform, mel-spectrogram, and MFCC visualizations.

Human-in-the-loop browser automation system with a multi-layer scheduling architecture. A FastAPI app deployed on Render (Docker) stays alive via UptimeRobot + GitHub Actions keepalive. At scheduled times, cron-job.org triggers the API, which reads email confirmations via IMAP and — upon approval — dispatches a Selenium (Chrome headless) agent to execute the browser workflow with randomized timing delays. Action results are reported back via Resend API.

Parallel web scraping and data intelligence system that monitors 6 Brazilian e-commerce competitors across 12 product categories, extracting 20+ attributes per product. Each retailer has a dedicated Selenium + BeautifulSoup4 scraper tailored to its DOM structure, with hybrid extraction (JSON-LD structured data + CSS selectors), dynamic scroll/pagination, and user-agent spoofing. A ThreadPoolExecutor runs all 6 scrapers concurrently, bulk-inserts results into SQL Server via SQLAlchemy + pyodbc (parameter-limit-aware batching), and passes the data through a 50+ rule normalization engine (colors, sizes, materials, attribute inference). Final output feeds a Power BI dashboard for pricing and product mix analysis.

Local data lakehouse replicating AWS architecture (S3 → Athena → MWAA) entirely on open-source tooling. Uses Terraform to provision MinIO buckets, Dagster to orchestrate a medallion pipeline (raw → validated → trusted) over StatsBomb open data, Great Expectations for data quality gates (coordinates, IDs), and DuckDB for stateless Parquet queries. Full Docker Compose setup with GitHub Actions CI/CD.

Windows desktop app that maps any USB/Bluetooth gamepad to keyboard, mouse, and macro sequences — no drivers required. Built with customtkinter (GUI), pygame (60 Hz joystick polling), and pyautogui/pynput (OS-level input simulation via Windows SendInput for Raw Input compatibility). Supports key combos, hold-while-pressed, timed macros, multi-step action sequences, analog stick → mouse, and JSON-based preset system. Distributed as a standalone .exe via PyInstaller.

• 🤖 RAG & LLM Orchestration: Building LangGraph-based pipelines with multi-query generation, LLM relevance validation, and vector stores (pgvector, Qdrant) for production economic analysis systems.
• 🧮 Vector Search & Embeddings: Designing production ETL pipelines for large text corpora — from noisy ingestion and NLP preprocessing to dense embedding storage and semantic retrieval.
• 🧠 Machine Learning: Applying classification, forecasting, and pattern recognition models across domains — from financial time series to audio feature engineering.
• ☁️ Cloud Architecture: Optimizing data pipelines and efficiency on AWS and Snowflake.

Whether it's about AI, Cloud Engineering, Automation, or Retro Gaming, I'm always happy to chat! 📬 LinkedIn | ✉️ restier2001@gmail.com