A forecast becomes an instruction on one specific farm.
Everything below exists to close the distance between a county-scale weather bulletin and a woman standing in a valley field deciding whether to harvest today.
National forecasts stop at the county line.
Kenya Meteorological Department forecasts are authoritative, and they are the foundation we build on rather than something we replace. But they are issued at county scale, in technical English, and they arrive with no delivery channel into a low-literacy, feature-phone household.
The data thins out exactly where the farms are. Ground-level weather observation across sub-Saharan Africa runs roughly eight times below the minimum station density the World Meteorological Organization recommends. That sparseness is what inflates basis risk, the gap between what an insurance index pays out and what a farmer actually lost, and basis risk is the single most-cited reason smallholders distrust index products.
So the problem is not that the forecast is wrong. It is that a farmer on a valley plot cannot learn from it whether her field floods on Thursday. We work entirely inside that gap.
Deliberately boring, on purpose.
Our alert engine launches with the least artificial intelligence in the whole platform, and that is a design decision rather than a limitation of what we can build.
Version one is a deterministic rules-and-thresholds engine running over open meteorological data, crossed against each registered plot's location, terrain and crop profile. Deterministic means every alert can be explained: to a farmer, to a champion, to a county official, to an underwriter. It also means the system fails conservatively rather than creatively.
That matters because our own risk register is explicit that false alarms are the fastest way to destroy this business. A discredited early-warning system is more dangerous than no system at all, because it teaches people to ignore the warning that finally matters. So machine learning earns its way in only as verified ground truth accumulates: downscaling forecasts to ward level, fusing terrain and satellite data into plot-level flood probability, and learning from real outcomes which thresholds actually predicted events.
At pilot scale, every single alert passes human review before it is sent.
Her language. Her reading level. One action.
An alert that a farmer cannot read, or can read but cannot act on, is not an alert. It is noise with a delivery receipt.
So every message is written in Dholuo or Kiswahili at a Standard-6 reading level, and carries exactly one protective instruction: harvest mature cassava now, move stored seed to higher ground. Never a probability, never a forecast summary, never a paragraph.
Dholuo is a low-resource language that no commercial agritech platform serves well, which is precisely why we treat it as core engineering rather than a localisation afterthought. Large language models carry the translation and parse farmers' free-text and voice-note replies back into structured records, and every message class is sample-reviewed by native speakers before and during deployment.
What we deliberately do not do
- We never let a model freely generate agronomic advice. Crop-disease recommendations are constrained to a validated library, and low-confidence cases route to a human.
- We never phrase an alert as a command that overrides what a farmer can see. Champions are trained to teach farmers to weigh our alert alongside their own seasonal knowledge, which is a resilient capability built over generations.
- We never promise a payout, a yield or an income outcome. We claim the mechanism and the measured result, and nothing else.
Built for a KSh 1,500 phone, not a flagship.
The primary channel. Works on every handset in the village.
Registration on 2G, in under five minutes, with a champion sitting beside her.
Voice-note replies so low literacy is never a barrier to being heard.
A weekly Dholuo programme carries warnings past the edge of our registry.
A farmer-facing mobile app would contradict everything above, so it is not on our roadmap. WhatsApp comes later, for the households that already have a smartphone, and never as the front door.
One number we cannot fake.
Most climate services report how many messages they sent. That is an output, and it is almost meaningless: a message delivered to a phone nobody read changed nothing at all.
Our north-star metric is Farmer-Acted Alerts. It counts the farmers per month who take a verified protective action in response to one of our warnings, confirmed by a champion's phone call and cross-checked by independent survey sampling.
One auditable number proves the entire chain end to end. The alert was delivered, so the infrastructure works. It was understood, so the language design works. It was trusted, so the education works. It was accurate, so the data works. And it was acted on, so behaviour actually changed. If any link fails, the number falls, and we cannot hide that from a funder, an underwriter or the community itself.
Verification is sample-audited independently of the champion who did the verifying, because concentrating onboarding, support and confirmation in a small group of trusted individuals creates exactly the incentive to inflate a number.
Want the technical detail?
We will walk any serious partner through the alert logic, the verification protocol and the data-governance design, in as much depth as you want.
Shamba Salama