East Africa · Agriculture Live · Deployed

AgroFutures

Agricultural advisory and regional pest & disease early-warning for smallholder farmers in Kenya, delivered over any phone, in the farmer's own language.

Farm Advisory

The individual farmer's tool.

Purely informational: sign up, get AI advisory and tips on your own plot. Nothing is verified or reported anywhere. Smartphone gets the web dashboard, WhatsApp, and Telegram; any other phone gets the same advisory over USSD and voice callback, in the farmer's own language.

Where It Runs
Channels
Any phone the farmer already owns.

2G USSD for feature phones. Voice IVR with local-language text-to-speech for callers who'd rather listen than read. And, as of today, Telegram and WhatsApp for the growing number of farmers on a smartphone.

One advisory engine, met on whatever channel the farmer already uses. No app to install, no data plan required for the USSD and voice paths.

USSD · Voice IVR · Telegram · WhatsApp
Reach
2G covers 98% of Kenya's population: wider than 4G.

Because the advisory runs on the 2G layer, it reaches any phone anywhere there's a signal, with no smartphone and no data plan needed. That coverage figure is the addressable footprint, not a registered-user count. (Coverage: Safaricom.)

98% 2G population coverage · Kenya
Featured · Live behaviour

It reasons against the farm, and hedges when the photo doesn't fit.

On Telegram and WhatsApp, a farmer can send a photo of a sick plant. Most tools would return a raw image-classifier score and stop there. AgroFutures doesn't.

It checks the photo against what it already knows about that farm: growth stage, vapour-pressure deficit, recent weather, soil. When the classifier's read doesn't fit those conditions, it visibly revises or hedges its own answer and asks for more evidence, rather than forcing a false diagnosis in either direction.

That's the reasoning discipline the whole platform is built on: confirm, discard, or flag. Never silently trust a score, never silently reject an out-of-pattern signal. It's a real, demonstrable behaviour, shown in the recording, not a claim.

Telegram / WhatsApp photo diagnosis.

See It Running

USSD advisory in Dholuo (Luo).

USSD farmer registration in Kikuyu.

Marine pest/ocean alert broadcast.

USSD advisory flow, feature phone.

Voice callback advisory (IVR).

The Advisory Pipeline

From scattered reports to a warning a farmer can act on.

01 · Cluster

Text reports from many farmers cluster into an outbreak signal before any single farmer could see the pattern.

02 · Confirm or reframe

Where a photo exists, it confirms or reframes the signal using farm-specific context (growth stage, soil, weather) rather than trusting a raw classifier score.

03 · Resource & warn

Confirmed evidence generates input resourcing and preventive warnings, including for farms that never submitted a photo themselves.

04 · Return over the same channel

Alerts go back to farmers over the channel they already use: USSD, a voice callback in their language, or a message thread.

Language Is an Access Problem

Seven thousand languages are spoken on earth; the major AI platforms were trained on a fraction. We use Khaya AI (GhanaNLP) for African-language TTS and are integrating PazaBench ASR models from Microsoft Research Africa for voice input in Dholuo, Kalenjin, Kikuyu, Maasai, and Somali. Language is not a localisation problem. It is an access problem.

Why it matters A farmer who receives a mistranslated pest advisory doesn't just miss the information. They may act on the wrong one. Generic translation APIs have, in our testing, misidentified regional languages and mistranslated local crop-variety names into nonsense.
Verification note Specific named-vendor examples are re-verified before publication, since low-resource-language model quality shifts month to month.
Regional Portfolio Manager

The officer's tool.

Kenya's Ethics and Anti-Corruption Commission found that ward agricultural officers tasked with verifying farmers' details weren't facilitated with transport or ICT tools. Farmers ended up financially supporting officers just to get a visit, and cartels hoarded input vouchers while genuine smallholders got locked out. That's not a story about bad officers. It's a resource constraint: one officer covering 1,000+ farmers with no vehicle and no targeting data will burn a week visiting the wrong ones. Regional Portfolio Manager is built against that constraint directly, offline-first and syncing on reconnect.

Precision Targeting

Scattered USSD and photo reports cluster into one confirmed ward-level signal, so an officer knows where to go before they leave. One visit becomes a ward-wide broadcast instead of a blind farm-by-farm trip: deployment that fits the 1,000+:1 ratio instead of fighting it.

One-Plot, One-Marker Registration

GPS registration is locked to one marker per physical plot. That single rule does two jobs at once: it's what makes a plot's EUDR deforestation-screening result actually mean something (the coordinates checked are the coordinates farmed), and it's what lets each plot draw down its input voucher once instead of twice. Market access and voucher accuracy both run through the same clean coordinate.

Offline Integrity, Future-Proofed

Every voucher approval, price update, dispatch, and payout writes to an append-only hash-chained ledger. Not a blockchain flex. Integrity for writes that originate offline and need to stay tamper-evident by the time they reach the server, and a ledger that's already recording what future seed-provenance and chain-of-custody rules will require, before those rules exist.

Human-Gated Advisory

Reports come in via USSD text and WhatsApp/Telegram photo; AI drafts an advisory grounded in 35+ live data points per farm and self-rates its confidence, but nothing broadcasts until an officer or agronomist approves, edits, or replaces it.

regional.agro-futures.com ↗
AgroFutures

The intelligence is frontier. The phone is $15.

Microsite ↗ GitHub ↗ Work with us →