Senior Data Scientist building AI Agents & LLM systems that run in production — at Braskem, Bradesco, B3 (Brazil's Stock Exchange) and beyond
6+ years turning messy business problems — fraud, AML/CFT, churn, credit risk — into ML systems that ship. These days that means AI Agents and LLM applications on Azure AI Foundry; before that it meant fraud models running on Databricks catching real money in real time.
- Building multi-agent LLM systems and intelligent document processing at Braskem
- Exploring: agent evaluation/observability, RAG architectures for regulated industries
- Writing about ML fundamentals (metrics, cross-validation, bias-variance) on Medium
| Repo | What it shows |
|---|---|
| brazil-municipal-gdp-regression | Regression on real IBGE data (5,570 municipalities): EDA → baseline → 5-fold CV model comparison → Optuna Bayesian tuning → SHAP explainability. XGBoost hits R² 0.90 in CV, but SHAP traces the CV-vs-test RMSE gap straight back to a single mining-town outlier the model never learned to extrapolate to. Shipped as a Dockerized FastAPI service with Supabase-logged prediction history. Try it live → |
| fatal-accident-prediction-br | Imbalanced binary classification (13:1) on 73k real PRF traffic-accident records: EDA → baseline → 5-model CV comparison (LogReg, RF, XGBoost, CatBoost, LightGBM) → Optuna tuning + threshold sweep → SHAP explainability. Exposes the "93% accuracy, 0 fatalities detected" trap, then fixes it — F1 climbs from 0.00 (dummy) to 0.39, ROC-AUC 0.835. Shipped as a Dockerized FastAPI prediction service. |
| municipios-br-clustering | Unsupervised clustering on real IBGE data (5,570 municipalities): EDA → baseline KMeans → 7-algorithm comparison (KMeans, Agglomerative, DBSCAN, HDBSCAN, OPTICS, KModes, KPrototypes) → final model shipped as a public API. Rediscovers Brazil's Southeast/Northeast economic divide with zero labels — silhouette 0.41. Try it live → |
| anp-fuel-price-anomaly-detection | Unsupervised anomaly detection on 52k real ANP fuel-price records across all 27 Brazilian states: EDA → baseline → 4-algorithm comparison → synthetic-anomaly evaluation (no real labels exist, so anomalies are injected to measure it). Exposes LocalOutlierFactor flagging zero anomalies despite a 0.91 ROC-AUC; winner is OneClassSVM, F1 0.68. Shipped as a live map, auto-updated monthly by GitHub Actions. Try it live → |
Senior Data Scientist — Braskem (current)
AI Agents, multi-agent systems, LLMs, prompt engineering, and intelligent document processing on Azure AI Foundry — moving enterprise workflows from manual to automated.
Data Scientist — Banco Bradesco
Built a LightGBM fraud-detection model on large-scale transaction data: benchmarked candidates with Databricks AutoML, handled severe class imbalance through sampling, and validated with time-based cross-validation to avoid look-ahead bias. Tuned with Optuna, tracked with MLflow, and orchestrated end-to-end through Databricks Jobs.
Shipped to production at an 80% fraud detection rate.
CRM Data Scientist — Banco Sofisa
Lead propensity models, customer segmentation, and recommendation systems feeding credit and CRM decisions.
Within 3 months of deployment, the propensity models drove R$6M+ in new credit risk originated for the bank
Data Scientist — B3 (Brazil's Stock Exchange)
Built econometric revenue-forecasting models (frequentist and Bayesian) for macroeconomic risk monitoring, and unsupervised fraud detection (KMeans, DBSCAN) to profile investors behind fraudulent trading activity — plus risk-control tooling that automated the monthly monitoring of funds trading assets they weren't authorized to hold.
DBSCAN isolated fraud cases into their own distinct cluster — turning a manual investigation into a repeatable detection signal.
Languages Python · SQL · R ML scikit-learn · XGBoost · LightGBM · CatBoost · TensorFlow · PyTorch GenAI / Agents GPT · Claude · Azure AI Foundry · RAG · Multi-Agent Systems · Prompt Engineering Data & MLOps Databricks · PySpark · Delta Lake · MLflow · Feature Engineering Cloud Microsoft Azure Other Git · Docker · Power BI
Instructor at FCCD and Universidade dos Dados — courses on Machine Learning, Statistics, and Python. Explaining a concept clearly to a room of students is a good forcing function for actually understanding it.
Electrical Engineering (UNESP) · Postgraduate in AI & Big Data (USP)
🇧🇷 Portuguese — Native · 🇺🇸 English — Professional working proficiency · 🇫🇷 French — Intermediate
📫 ejunior029@gmail.com — open to conversations about AI Agents, LLM applications, and ML in production.