TerraRisk SP

Brazilian municipal health managers have no way to know where to concentrate preventive resources against climate-sensitive disease — dengue, malaria, leptospirosis. The data that would let them prioritize exists, but it is fragmented across four public bases nobody joins at municipal level: climate (INMET / Copernicus), disease (DATASUS), land use (MapBiomas), population (IBGE). The existing alternatives were international consultancies at US$50K+/year or one-off academic studies.

I built TerraRisk — a production platform that integrates the fragmented public data into a single municipal-level dataset and turns it into a decision tool, delivered as a “serious game” workshop that walks officials from intuition to a data-driven ranking. I own the platform and the data engineering; my co-founder owns the statistical model.

Multi-source geospatial ETL

I built the pipeline that ingests and harmonizes heterogeneous sources — Copernicus CAMS/ERA5 (NetCDF), CHIRPS (GeoTIFF), DATASUS, IBGE census, MapBiomas — via spatial joins against IBGE municipal shapefiles (EPSG:4326) and zonal aggregation, producing an integrated dataset of 645 municipalities × 104+ variables (CSV + GeoJSON), published FAIR-compliant with a Zenodo DOI.

The platform

• Frontend: Next.js 14 (App Router), Leaflet interactive maps, dynamically generated bivariate maps, gamified credit system, 4-phase workshop flow, i18n (PT/ES/EN).
• Backend: FastAPI with 7 routers serving the dataset and powering the workshop.
• Operationalizing the science: ~98 Python scripts that turn a peer-reviewed risk model into executable analysis — geospatial processing (GeoPandas, Shapely), the integrated dataset build, policy-gap computation. I make the science run, reproduce and ship.

Satellite data engineering

On the Air Pollution sub-project I extracted Copernicus CAMS/EAC4 (surface O₃, NO₂) + ERA5 across 4 batches with automated QA (13–18 checks per batch) — nulls, ranges, duplicates, cross-source matching — processing 1,419 of 1,476 networks (96.1%) and 218,000+ observations. Paid contract.

Results

• 645 municipalities (100% of São Paulo state); 104+ integrated variables, FAIR dataset with a Zenodo DOI.
• 1,419 / 1,476 satellite networks processed; 218K+ observations.
• Validated in a state-government workshop (SEMIL-USP, ~50 participants); presented to CEMADEN.
• Live in production, 4 containers on a ~$10/mo VPS, CI/CD.

This is the data half of modern AI engineering: ingesting messy, heterogeneous, real-world sources; harmonizing them spatially at scale; building reproducible pipelines; and shipping the result as a live product institutions actually use. Paired with Vindex, it covers the full span — data engineering through applied LLM systems.