Answer farmers' questions on WhatsApp from the government farming manual
Opportunity International's Ulangizi AI lets smallholder farmers in Malawi ask farming questions by voice, text or photo on WhatsApp, in Chichewa or English, and answers mainly from the government's own farming manual, citing its source.
Built withChatGPT / OpenAI models
Try it withClaudeChatGPT / OpenAI modelsGeminiGrok
Deployed; available in Malawi since February 2024 (per NVIDIA, Sep 2024) and still in use in October 2025 per Rest of World, with a sister service, FarmerAI, in Kenya, Opportunity International, with Gooey.AI and Malawi's Ministry of Agriculture (source manual)
Why it matters
About three in four Malawians farm, and more than 80 percent live in rural or remote areas, but Malawi's Ministry of Agriculture says there is about one extension officer for every 3,000 farmers. One extension worker told Rest of World she oversees 1,743 farmers and rarely sees a group more than once a month. Around 30 percent of adults do not read or write, so the service takes voice notes as well as text and photos, and trained Farmer Support Agents relay questions for farmers who have no smartphone. NVIDIA, a technology partner, reports that it cut the time to get farming guidance from days to seconds, and Opportunity International says the first three months brought more than 4,000 questions in English and Chichewa. GiveDirectly has helped 2,800 farmers get phones with the app.
How it works
Farmers message the service on WhatsApp (Gooey.AI also lists SMS). A voice note is transcribed and translated into English with speech models such as Meta's MMS or a fine-tuned version of OpenAI's Whisper, and photos of crops or animals go to GPT-4o. The question is then matched against a vector database built from the Ministry of Agriculture's 450-page production manual, and the language model writes an answer from the passages it finds, in English or Chichewa, as text or audio. Each answer includes citations and links to the relevant source. Opportunity International built it with Gooey.AI, which runs the workflow and regularly swaps in newer models; the service runs on NVIDIA GPUs in Microsoft Azure. Farmer Support Agents, who receive a smartphone and training, use it with the farmers they serve. The same approach runs in Kenya as FarmerAI, with Safaricom, in English and Swahili.
Who it's for
Agricultural extension services, farmers' cooperatives, agri-lenders and NGOs that advise smallholder farmers, especially where officers are few and many people prefer voice to text.
The prompt
You are helping me design a WhatsApp farm adviser that answers farmers only from an approved farming manual and says where each answer comes from. Do not ask me for the names, phone numbers or locations of any real farmer. Our organisation and area: [for example, a farmers' cooperative or an agricultural extension service in one district] Our approved source: [for example, the national good agricultural practices manual, with its title and year] Languages and how people will ask: [for example, a local language and English, mostly by voice note] 1. List the ten questions farmers here are most likely to ask in a season (planting, pests, diseases, fertiliser, storage, selling), and for each, which section of our source should answer it. 2. Write the rules the adviser must follow: answer only from the approved source, quote the section it used, say plainly when the source does not cover the question, and never recommend a pesticide, dose or medicine that the source does not name. 3. Draft three sample answers to voice-note questions, each under 80 words, in plain words a farmer can act on, with the source section at the end. 4. List the questions that must go to a person (for example suspected animal disease outbreaks, pesticide poisoning, or anything about money or loans), and write the exact handover message. 5. Design how a field agent helps a farmer who has no smartphone: what the agent asks, what they send, and how they pass the answer back. 6. Draft a plain notice telling farmers that answers come from AI, what is stored, and how to reach a person. Do not invent manual sections, products, prices or phone numbers; leave placeholders. Mark every rule that an agronomist or extension officer must sign off before use.
Open inClaude (opens in a new tab)ChatGPT (opens in a new tab)Gemini (opens in a new tab)Grok (opens in a new tab)
Sources
The NVIDIA blog is written by a technology partner and draws on Opportunity International's own account; the Gooey.AI page is the builder's. Rest of World reported independently, interviewing farmers, a field agent, an extension officer and a university vice-chancellor. No independent evaluation of outcomes was found.
1. NVIDIA Technical Blog: "Since February, farmers in the East African nation of Malawi have had access"; "more than 4,000 questions"; "a 450-page book"; "citations and hyperlinks". AI Chatbot Delivers Multilingual Support to African Farmers (opens in a new tab)
2. Gooey.AI (builder): "AI-powered chatbots that deliver agricultural advice on demand"; "supported 1,800 farmers, generating over 17,000 user messages". Ulangizi AI (Farmer AI project page) (opens in a new tab)
3. Rest of World: "one extension officer available per 3,000 farmers"; "Ulangizi can sometimes generate false answers"; "2,800 farmers in obtaining phones". Malawi’s new farmhand: AI that speaks the local language (opens in a new tab)