AgroFutures

Three climate zones.
One AI engine.
Last mile included.

A terrain-aware agricultural intelligence platform active across Mediterranean, tropical highland, and tropical coastal zones. It delivers per-plot AI advisory (text, spoken voice notes, and photo diagnosis) from a zero-data USSD line to Telegram and WhatsApp, integrating OSRM cold-chain routing with ledgered 2G farm-gate transactions. The same engine reaches a Sardinian olive grove and a Kiambu maize plot alike.

USSD · 2G
SMS
Voice notes
WhatsApp · Telegram
Photo AI

See it work.

Not a mockup. The real bot, on a real phone: a leaf photographed, a disease named, the answer spoken back in the farmer's language.

Photograph a pest. Get the answer.
A leaf photo → Fall Armyworm, 95%, fused with this plot's humidity & stage, in Dholuo.
The advice, spoken.
Every answer as a voice note in the farmer's own language. Literacy is never the barrier.
Ask on any phone.
A free-text question on a 2G line → an AI answer by SMS, grounded in your plot.
PROBLEM 01: REGISTRATION

The EU's deforestation regulation requires plot-level geolocation and farmer identity for every agricultural product entering European markets. Kenya could lose an estimated KES 90 billion in export earnings over five years for non-compliance, but only 30% of coffee farms have been geo-mapped so far. Farmers have limited access to internet and digital tools, the regulation's language is too technical for most to understand, and government outreach has barely reached the rural level. (Source: IPS News, Jan 2026; Kenya Agriculture & Food Authority)

Solution built. Instead of pen-and-paper registration at schools, farmers sign up once via 2G USSD, creating a persistent digital record that feeds GPS plot verification and EUDR deforestation checks per farm.

PROBLEM 02: INTELLIGENCE DELIVERY

3.1 billion people live under a mobile signal and still can't access the internet. The GSMA identifies why: device cost, data cost, limited digital literacy, no local content. Every AI advisory platform ever built requires all four problems solved first. (Source: GSMA, cited Forbes February 2026)

Solved. USSD, SMS and voice callback on 2G, and on Telegram and WhatsApp, free-text AI chat, spoken voice notes, and photo diagnosis. One AI brain, every channel, delivered in native languages via Khaya AI. Zero data cost on the phone that can only make a call; richer for the phone that can send a picture.

Western AI platforms are built for 3,000-acre commercial farms at $10,000+ a year. Every app-based alternative, however cheap, starts with a $75–90 smartphone that consumes 26% of the average monthly income in Sub-Saharan Africa, plus a data plan. This platform costs about one US cent per advisory and runs on any phone that can make a call. (Smartphone cost: Kenya retail 2026. Income figure: GSMA)
PROBLEM 03: DATA DESERTS

Africa holds 60% of the world's uncultivated arable land but can't plan, lend, or insure because the agricultural data doesn't exist.

Solution built. Unstructured 2G USSD reports (pest sightings, flood events, crop stages, yield estimates) are converted into clean database records, documenting regional productivity over time. The farmers generate the data on phones they already own. The platform doesn't just deliver intelligence; it builds the structured dataset that planning, lending, and insurance require.

PROBLEM 04: VISIBILITY

A farmer knows what they planted and roughly what they harvested. They don't know their yield gap, their soil penalty, what heat stress is costing them, or what a different crop would earn on the same plot. The data exists in satellites, weather stations, and FAO reports, but it's never been assembled per plot and delivered to the person farming it.

Solution built. Terrain, soil type, yield gap, drought exposure, and crop pivot economics, assembled per farm from satellite, weather, and market data. Most of these farmers have never seen their own plot from above; satellite imagery is processed on the server and distilled into the numbers that matter (NDVI, slope, soil type) so that first view of their land arrives on a 2G screen instead of a smartphone they don't have. The same JSON feed powers the web dashboard and the USSD line.

PROBLEM 05: COMPREHENSION

Agricultural extension has always assumed the farmer already speaks the vocabulary of soil science and plant pathology. Most don't. "Vegetative stage." "Systemic pesticide." "Frass." No amount of reach or speed fixes advice that arrives in words the recipient can't parse.

Solved. Glossary lookup from the same shortcode, same 5 languages. A term they don't recognize gets a plain-language definition: Wikipedia where one exists, AI-generated where it doesn't, cached and ready the next time they dial in.

One platform. One shortcode. All five problems.

One farmer. One session. The whole loop.

Tonny dials *384#
Registered in 30 seconds Problem 1
AI advisory in Dholuo: Fall Armyworm risk HIGH Problem 2
Types back: "why are my leaves yellow?"; AI answers by SMS, in Dholuo, grounded in his plot Problem 2
Doesn't recognize "frass"; looks it up, gets a definition in Dholuo Problem 5
Reports: pest on maize, whole field, severity 3 Problem 3
Cluster alert fires. Extension officer dispatched. Problem 3
Checks land info: NDVI, slope, soil type; sees his own plot from above for the first time Problem 4

Every step on a 2G signal. Every report makes the next advisory smarter for every farmer in the zone.

A crop advisory in English helps some farmers.
In their mother tongue, it reaches everyone.

English · Kiswahili · Dholuo · Kikuyu · Kimeru
Khaya AI (GhanaNLP) handles all African-language output. Current: English, Kiswahili, Dholuo, Kikuyu, Kimeru. Expanding to 12+ languages, actively pursuing grant funding to scale. Native Kikuyu speaker validates agricultural terminology. Moved from Gooey/Google Translate to Khaya after discovering critical domain failures (coffee varieties translated as personal names in Kikuyu). And Khaya doesn't just translate; it speaks: every answer comes back as a spoken voice note in the farmer's own language, so a farmer who cannot read still hears the advice. Correct language and voice aren't polish on the product; for the person on the far end of the line, they are the product.

Inside the AI Engine

Every advisory starts with a dense agronomic brief assembled from live farm data. DeepSeek processes it. The result is distilled to fit inside a 2G USSD session, using less data than a single webpage image.

Inputs: Per Advisory
  • YTD cumulative rainfall vs ideal, month by month
  • NDVI: satellite canopy health score
  • VPD (vapor pressure deficit): transpiration stress signal
  • Soil moisture saturation %
  • Slope + erosion risk calculation
  • Biological pest profile: optimal temps, lifecycle days, outbreak thresholds, treatment window
  • Monthly temperature + humidity history
  • Season yield performance scores
Output · USSD · 2G
  • 182 characters per screen
  • 3–5 screens per full advisory
  • ~1–2KB total session data
  • Pre-processed on the server. The phone never talks to the AI directly. It talks to a shortcode.
14
Crop pest profiles
5
Advisory fields per crop

A second engine reads pictures. A farmer photographs a sick leaf or a pest; a multimodal vision model (Gemini Flash) identifies the disease or insect, rates its own confidence honestly, offers a differential, and (this is the point) fuses the finding with that exact plot's conditions: its VPD, soil, water deficit, growth stage, known local pest pressure. Not "here is what Fall Armyworm is," but "this is Fall Armyworm, on your maize, at this stage, and here is why your humidity makes it worse." At roughly one hundredth of a US cent per diagnosis.

Every confident diagnosis enters a review queue, not the answer library. A local expert confirms, corrects, or rejects it, and only human-verified cases feed back into the model. The AI is fast; the expert is the truth. That is how the dataset stays clean, and how it earns the trust the advice depends on.

Three Climate Zones. One Engine.

The platform was built as a full web dashboard: four climate zones, each with satellite data, yield modelling, drought scenarios, and crop pivots. Then came the harder problem: the same intelligence was completely unreachable to the smallholders who needed it most. USSD was the answer.

Zone 01 · Mediterranean
Alghero, Sardinia

Olive groves, tree crops, wine. Satellite NDVI, rainfall/yield overlays, drought scenario modelling, crop pivot analysis. Bombarde Farms is the live deployment. Six years of data.

Zone 02 · Tropical Highlands
Kiambu County, Kenya

Maize, mixed smallholder farming at elevation. Gachororo Community Co-Op (Kiambu) is the live client for AI advisory. Full dashboard plus USSD and voice delivery for farmers without smartphones or data.

Zone 03 · Tropical Coastal
East African Coast

Near-ocean farming: coconut, cashew, cassava, mixed coastal crops. Humidity, salinity, and wind-risk modelling. Soil profiles tuned for coastal conditions.

The model advising a Sardinian olive grower on irrigation timing is the same model that tells a Kiambu maize farmer whether Fall Armyworm risk is high, delivered over a 2G signal, in Kikuyu, for one cent. Same engine. Every phone. Every zone.

The Platform

Farmer Layer
  • USSD Advisory
    *384# · 5 languages · zero data · any phone
  • USSD Ask-a-Question
    free-text question → AI answer by SMS, farm-fused, in-language
  • Telegram & WhatsApp Bots
    advisory · free-text chat · crop select · same brain, richer screen
  • Photo Diagnosis
    snap a leaf/pest → AI ID + plot-fused advice New
  • Voice Notes
    every answer spoken in the farmer's language (Khaya TTS): the literacy layer
  • Voice IVR
    outbound calls · Khaya AI vernacular TTS
  • SMS Field Reports
    Data Desert Killer
  • USSD Registration
    identity · location · farm size · language
Intelligence Layer
  • DeepSeek Advisory Engine
    14 crop profiles · 6-year validated model · Sardinia to Central Highlands
  • Gemini Flash Vision
    photo → disease/pest ID · calibrated confidence · differential · fused with the plot's live conditions
  • Expert-Verified Evidence Store
    photo diagnoses ≥50% confidence queued for expert review; confirmed cases feed back into the engine; AI fast, human true
  • Khaya AI Translation & Voice
    GhanaNLP · Dholuo · Kikuyu · Kimeru · text + spoken · expanding

What The Platform Actually Does

This is NOT just "crop advisory in 4 languages." It is a terrain-aware, financially-modeled, drought-scenario-tested, multi-pivot agricultural decision engine. Here's everything it does:

Soil & Terrain Intelligence
  • Terrain Intelligence Slope analysis with microzone recommendations: "North-facing, 12° slope, 1526m elevation, runs 2°C hotter than neighbouring flat plots" Bottom third: highest moisture, best for maize and water-hungry crops Top third: dries first, most wind, plant drought-tolerant crops here VPD (vapor pressure deficit) stress monitoring Irrigation timing against rainfall forecast: "Next rain: Fri 06:00 PM. Hold irrigation until after" NDVI crop health tracking: "+22.9% since last satellite pass"
  • Soil Behaviour Soil type profiling (Ferralsols, Calcisols, etc.) with specific management advice (pH, drainage, fertility, compaction risk) Crop-soil matching: "This soil rewards Maize: Cassava, maize, coconut, mango" Organic matter and nutrient management recommendations
  • Yield Gap Analysis Current achievement vs land maximum (e.g., 68.8% = 1.7t/ha, gap = 0.8t/ha) Revenue gap in KSh: "KSh 156k uncaptured" Prioritized bottleneck identification: "Fix this first: Rain timing mismatch" Potential gain per fix: "+KSh 40k if resolved"
Risk Modeling & Crop Pivots
  • Drought Scenario Modeling Stress-tests farm economics before drought hits Shows transition flow: Normal Revenue → Drought Revenue → Drought Profit → New Break-even Viability assessment: "Maize remains viable under drought. Break-even rises to 1.3t/ha, still achievable."
  • Value Chain / Product Options Same crop, different processing paths Example comparative analysis: Raw Maize (KSh 344k) vs Silage/Livestock Feed (KSh 308k + KSh 71k = 1.3x revenue) Economic thesis: "Same land · Same water · Same crop · 1.3x revenue multiplier"
  • Crop Pivot Recommendations (6 Categories, 10+ Crops) Each recommended pivot includes profit projection, water savings, yield estimate, price, and cited source.
    Resilient Food: Banana (TC): KSh 699k profit, flood-adapted; Sorghum: KSh 517k profit, waterlogging tolerant; Sweet Potato: KSh 663k profit, fast-maturing; Taro (Arrowroot): KSh 596k profit, naturally flood-tolerant.

    Bio-Energy: Napier Grass (Biogas): KSh 598k profit, dual revenue: feed + biogas; Water Hyacinth (Biogas): KSh 329k profit, zero water cost, circular economy.

    Textile: Kenaf: KSh 585k profit, industrial fiber demand; Sisal: KSh 398k profit, Kenya top-3 world producer.

    Pharma: Ginger: KSh 542k profit, good intercrop with banana; Turmeric: KSh 617k profit, organic export demand.

    Sources cited: HCDA Kenya, FAO, KEPHIS, KALRO, Kenya Sisal Board, KARI, NEMA
AI Engines & Crowd Intelligence
  • AI Advisory Engine DeepSeek via Cloudflare Worker proxy 14 crop pest profiles 5 advisory categories per crop: Headline, Water, Pest, Outlook, Microclimate Farmer pest/weather reports captured as ground truth for the advisory engine 6-year validated model across two hemispheres (Sardinia olives → Kenya maize)
  • Photo Diagnosis Engine (Gemini Flash) One multimodal call: identifies plant disease OR insect pest, no separate services Self-rated, calibrated confidence, honestly lower when the image is ambiguous Differential diagnosis (a ranked shortlist), like a real agronomist Fused with the plot's live conditions: VPD, soil, water deficit, growth stage, known pest risk Answer returned in the farmer's language, as text and a spoken voice note Domain-agnostic: reads the farm's climate zone + coordinates, not a hardcoded country ≈ one hundredth of a US cent per diagnosis
  • Expert Review & Verified Fact Base Every diagnosis ≥50% confidence lands in a review queue (Cloudflare D1), status "pending" A local expert confirms, corrects, or rejects, with the AI's candidate + differential as a head start Human-verified cases feed back into the advisory engine's fact base. The AI is fast; the expert is the truth Since 2G/USSD phones can't send photos, the minority with smartphone cameras builds the verified corpus that sharpens advice for everyone, including farmers on basic phones
  • Pest Cluster Detection (Crowdsourced Early Warning) Structured farmer reports via USSD: pest type, crop, severity 1 report = logged 2 reports same ward within 24hrs = officer SMS alert 3+ reports = county-level alert Cross-ward cluster = governor's office notified Clusters get a human visit — officer, expert, or co-op leader confirms on the ground; verification targets clusters, not every single report
  • Voice IVR Outbound calls via Africa's Talking TTS in English/Swahili native; Dholuo/Kikuyu/Kimeru via Khaya AI translate → TTS Interactive GetDigits menu for drill-down
Every interaction generates data. Pest reports, crop stages, harvest records, created by farmers on phones they already own. The platform doesn't just deliver intelligence. It builds the agricultural dataset Africa doesn't have.

Who It Serves

Farmers
AI advisory on any phone in any of 5 languages. USSD registration. SMS field reports. Outbound voice callbacks.
Extension Officers & Agronomists
Photo diagnoses arrive pre-triaged with an AI candidate and a differential: confirm, correct, or reject in seconds. Their verdicts become the verified knowledge base the whole system learns from.
Researchers & NGOs
Growing field report corpus. A live agricultural dataset built by farmers themselves.
Insurance & Parametric
Yield forecasting. Resilience scoring. Climate risk modelling. The field report corpus is the actuarial dataset that doesn't exist anywhere else.