Case study · embedded senior data scientist · Feb 2024 to Jan 2026
Climate pipelines at Raincoat, a parametric insurtech.
- Client
- Raincoat LLC, parametric insurance for natural catastrophes
- Markets
- LatAm, the Caribbean, and Africa
- Role
- Satellite and reanalysis data layer, new perils, dashboards
- Length
- 2 years, remote
The client
Raincoat builds parametric insurance products for natural catastrophe risks: policies that pay out automatically when predefined climate or weather thresholds are triggered, across emerging markets in LatAm, the Caribbean and beyond.
When I joined in February 2024, the company had already shipped three perils to production: Precipitation, Hurricane Wind and Earthquake, with Drought in early development. The codebase had grown peril by peril, and coverage was concentrated in the Caribbean and Colombia.
The brief
Peril ceiling. Sales conversations kept surfacing needs the existing perils did not cover: drought intensity, agricultural water stress, heat. New perils took months, because each one lived in its own block of code.
Geographic ceiling. The team was extending from the Caribbean and Colombia toward the rest of South America and Africa. Adding a country was a config change; the real cost was how much data the system processed per client.
My role was not to refactor the system. It was to plug in as a senior climate data scientist, extend coverage, ship new perils, and start showing what a more efficient architecture could look like.
What I contributed
Data acquisition, two new perils, three dashboards.
Data acquisition
Every peril except Hurricane Wind
Ingestion logic for CHIRPS, ERA5-Land, IMERG, Copernicus CDS and WEkEO, with automated refresh cycles and outputs that downstream pipelines could rely on. I also generated country-level maps for the sales team, which they used before contracts were signed.
New perils
Water Balance to production, Heat Stress ready
Water Balance was adapted from the Precipitation pipeline and shipped with CI/CD and Docker on Kubernetes. Heat Stress turned a research script into a production-ready pipeline; it was ready at contract end but not deployed during my time.
Dashboards
Three Plotly Dash tools used daily
Precipitation, Water Balance and Drought, used every day by the risk and underwriting teams. They were the interface between climate models and underwriting decisions.
Maintenance
Drought from early stage to stable
I maintained Precipitation and Drought for two years: calibration checks, geographic extensions, dataset updates. Drought became a stable peril in production across all target countries.
What changed in two years
Same team size, measurably wider coverage.
| When I joined · Feb 2024 | When I left · Jan 2026 | |
|---|---|---|
| Perils in production | 3 (Precipitation, Hurricane Wind, Earthquake), Drought early-stage | 4 (+ Water Balance), Heat Stress pipeline ready |
| Dashboards I built | 0 | 3 (Precipitation, Water Balance, Drought) |
| Coverage | Caribbean + Colombia | Most of South America + 4 African countries |
| Sales maps | Manual, on request | Country-level maps generated systematically |
What didn’t change: the perils already in production kept shipping reliably. No regression on Hurricane Wind or Earthquake during the period.
Key insight · the unfinished part
Five clients in one country meant five times the same downloads.
The architecture was monolithic per client and peril, so identical CHIRPS downloads and calibration runs were repeated for every client. Toward the end of my contract I sketched the alternative: a shared data pool per geography, computed once and refreshed on schedule, wrapped in reusable modules with standard inputs and outputs. New clients in covered geographies become near free to onboard, and compute scales with geographies, not clients.
CHIRPS v2 production ends after December 2026; new rainfall work uses CHIRPS v3.
I did not implement that refactor; it would have been a multi-month engagement of its own, out of scope for my contract.
How the work ran
Fully embedded in Raincoat’s remote team, time and materials, with weekly check-ins, shared repositories and direct contact with technical leadership and the risk team. Work ran project by project. The overlap between European mornings and American afternoons allowed real-time work without blocking either side.
Stack
- Climate data
- CHIRPS · IMERG · ERA5-Land · Copernicus CDS · WEkEO · NetCDF · Xarray
- Engineering
- Python · Pandas · NumPy · scikit-learn · PostgreSQL · Docker · Kubernetes · CI/CD · GitHub Actions
- Frontend
- Plotly Dash
- Methods
- Time-series analysis · Calibration · Threshold modelling · Anomaly detection · Geospatial aggregation
Hitting the same ceilings?
Adding geographies, adding perils, or getting a trigger checked before it pays. Fifteen minutes is enough to see whether I can help.