Project Overview
Kenya's National Drought Management Authority already publishes real, government-defined drought data every month — a five-phase system, Normal through Alert, Alarm, and Emergency, covering all 23 Arid and Semi-Arid Lands counties. But it's a monitoring system, not a forecasting one: it reports what's already true. So every month, coordinators and NGO field officers repeat the same manual cycle — read a dense PDF bulletin, judge which counties look like they're drifting toward crisis, then decide whether to act, often only after the phase change is already official and the lead time to prepare is gone.
Pulsecast turns that lagging report into a leading warning. A per-county forecasting model projects each county's vegetation-condition trend 4–6 weeks ahead and flags the exact date it's likely to cross into a worse NDMA phase — ranked into one priority queue, explained in plain language, and traceable claim-by-claim back to its source bulletin. A live regional map, a Kiswahili translation layer, and a backtest panel comparing past forecasts against what NDMA later confirmed round out a system built to be trusted, not just believed.
Then Pulsecast closes the loop forecasting alone can't: the moment a county enters Alert, Alarm, or Emergency, a coordinator can draft a real aid distribution and generate a ready-to-broadcast alert — bilingual, SMS-ready, tuned toward what that county's livelihood actually needs.
Built for NDMA county drought coordinators and NGO field officers under time pressure — and, through the alerts they send, for the residents of the counties those alerts reach.
Solution Details
Pulsecast scrapes NDMA's bulletins live, parses per-county VCI3M/SPI values from the PDFs, and feeds them into a per-county AR(2) forecasting model that projects 4–6 weeks ahead and detects the exact date a county crosses into a worse phase. A priority score — severity × time-to-crossing × confidence — collapses 23 separate forecasts into one ranked queue, cross-checked against ICPAC's independent seasonal outlook rather than trusted alone, and backtested against NDMA's own later bulletins so the track record is visible on-screen, not just claimed.
An AI layer translates the raw forecast into plain language, with every sentence grounded and citation-linked back to the exact source value — livelihood-aware, so pastoralist counties get guidance food-only counties don't. Output is available in English and Kiswahili, and a choropleth regional map surfaces neighboring counties drifting toward the same crisis that a list view would hide.
Then, the newest layer: the moment a county's confirmed phase reaches Alert, Alarm, or Emergency, a Send Aid panel unlocks on that county's page. A coordinator enters aid type, date, time, and drop-off points; Pulsecast composes the outbound SMS in both languages, sized to real segment limits, with pastoralist zones auto-nudged toward livestock feed and water. That message is deliberately template-built, not LLM-generated — a humanitarian broadcast is exactly the kind of high-stakes output that shouldn't come from a model that can hallucinate. This build simulates the send; wiring it to a live gateway like Africa's Talking is the disclosed next step, pending a funded, opted-in contact list.
Log in or sign up for Devpost to join the conversation.