Data Scientist & ML Engineer — rigorous modeling, production ML and LLM systems.

I build models that are evaluated the way they will be used: temporal and spatial hold-outs instead of random splits, honest baselines, calibrated probabilities and results reported even when the model only ties the baseline. Then I ship them — infrastructure as code, CI with type checks and coverage gates, monitored endpoints and evaluated LLM agents.

📫 gabriel.dias09@icloud.com · LinkedIn

Probability of ≥ 0.5 mm of rain in the next 1, 3 and 6 hours from INMET stations, METAR and Open-Meteo. A single XGBoost model with the horizon as a feature, a strict temporal split (train ≤ 2023, validation 2024, test 2025) and persistence and NWP baselines, scored with Brier, PR-AUC, CSI/POD/FAR and reliability tables. Deployed with AWS CDK (SageMaker pipeline, model registry with manual approval, PSI drift monitoring and retraining alarms). Includes a LangGraph agent on Bedrock with guardrails, evaluated against a 36-question golden set. 281 tests, 80% coverage gate. Python XGBoost AWS CDK SageMaker LangGraph Bedrock mypy --strict

Season-level (July–June) forecasting of yellow fever spillover, modelled as a wave that propagates rather than as static habitat suitability. Prospective and spatio-temporal block validation, decision curves and decile calibration; the final model is a 12-variable logistic regression with one representative per mechanism. Climate scenarios from a 5-GCM × 4-SSP CMIP6 ensemble with resampling-based uncertainty, plus an interactive dashboard. Python scikit-learn SHAP / ALE CMIP6 GitHub Actions

LightGBM Poisson on H3 cells × day × hour compared against simple baselines, Poisson GLM with clustered standard errors, fixed effects for speed cameras and a before/after design with placebo tests for speed bumps, plus a Monte Carlo simulator and a bilingual Quarto paper. Live dashboard. Python LightGBM statsmodels H3 Quarto

FastAPI + PostgreSQL + React. The bot uses a provider-agnostic LLM client (Gemini/Groq) with fallback, streaming, Retry-After handling and a bounded tool loop, plus speech-to-text with faster-whisper. CI runs the test suite against SQLite and PostgreSQL 17. FastAPI PostgreSQL React LLM tool use faster-whisper

EfficientNet-B0 on log-Mel spectrograms with attention pooling, MixUp, SpecAugment, mixed precision, MLflow tracking and Grad-CAM, validated with the competition metric on recordist-grouped folds so that no recordist appears on both sides of a split. PyTorch audio MLflow

Segmentation and measurement of Mimivirus fibrils in TEM images (CLAHE, optional U-Net, OCR of the scale bar) with a desktop GUI. OpenCV PyTorch computer vision

• Modeling: scikit-learn, XGBoost, LightGBM, PyTorch, statsmodels; time-series and spatial validation, calibration, decision curves, SHAP/ALE, count models, quasi-experimental designs
• ML engineering: AWS CDK, SageMaker, MLflow, Docker, GitHub Actions, pytest, mypy, ruff, uv
• LLM systems: LangGraph, Amazon Bedrock, Gemini/Groq APIs, tool calling, guardrails, golden-set evaluation
• Data & backend: Python, SQL, PostgreSQL, FastAPI, pandas, geospatial (H3, rasters, Earth Engine)

Also: bioinformatics tools from my research background — SAMANTHEX_2.0 (Sanger consensus GUI) and viral-ngs-pipeline.