Open-source AI school-heat advisory & learning continuity for South Sudan — county-level forecasts instead of blanket school closures
# HeatBell
**Open-source AI school-heat advisory and learning continuity — built for South Sudan, designed for replication in any heat-exposed LDC.**
> **Status: working prototype.** HeatBell runs end-to-end today — offline on a
> bundled sample and live against Open-Meteo
> forecasts. It is **not deployed with an education authority**, its tier
> thresholds are pilot defaults that **must be calibrated with health and
> education officials**, and its enrollment figures are illustrative
> placeholders. It is advisory decision support, not an early-warning mandate:
> **the education authority decides.**
---
## The problem
South Sudan has closed **every school in the country** over extreme heat two
years in a row — about two weeks in March 2024
as temperatures headed for 45 °C, and two weeks again from February 2025
after children collapsed from heatstroke in Juba. World Weather Attribution
found that climate change made the February 2025 heat **"at least 2°C hotter
and at least 10 times more likely"** — this is now an every-year hazard, in
the country with one of the world's most fragile school systems.
A nationwide blanket closure is the bluntest possible instrument: it costs
millions of children weeks of learning in counties that were merely warm,
while giving the hottest counties no earlier warning than anyone else.
**HeatBell turns the blanket nationwide closure into a county-by-county
precision response** — computed days ahead from free public forecasts, ranked
transparently, and delivered in the formats South Sudan actually uses: a
decision brief for the ministry and education cluster, a radio script, and
SMS texts.
## What it does
```text
Open-Meteo daily forecast (free, no key) config/southsudan.yaml (12 county seats)
\ /
v v
NOAA heat index (county x day, °C)
|
risk tiers: normal / elevated / high / extreme
(+ escalation: 3+ consecutive hot days raise one tier)
|
school-week view + dete …