Generative AI / LLMs

AI Chatbot Delivers Multilingual Support to African Farmers

An African farmer using an AI-powered chatbot on a cell phone asking a farming related question.

Some of Africa’s most resource-constrained farmers are gaining access to on-demand, AI-powered advice through a multimodal chatbot that gives detailed recommendations about how to increase yields or fight common pests and crop diseases.

Since February, farmers in the East African nation of Malawi have had access to the chatbot, named UlangiziAI, through WhatsApp on mobile phones. The chatbot understands agricultural-related questions sent by text or voice notes, and can also recognize and interpret photographs.

The app responds to queries in either English or Chichewa, the native language of about half of the country’s 20M people. In Malawi, where around 30% of the population doesn’t read or write, the apps ability to respond in natural language is crucial. 

It was developed by NGO Opportunity International, which has worked in Malawi for 21 years. Because smartphone access remains low in Malawi, the NGO hired a constellation of local Malawians to work with rural farming communities. These farmer support agents receive training and a smartphone, and then help farmers access the UlangiziAI app (“Ulangizi” in Chichewa means “guidance” or “counseling”). 

Paul Essene, who helped develop the technology behind UlangiziAI, said the chatbot can help overcome endemic, structural challenges Malawian farmers routinely face. 

Most Malawians live on around $2 per day. More than 80% of the country lives in rural or remote areas, and around three in four Malawians work on or own farms. Government offices designed to offer farmers advice are inconsistently open. Even getting to these offices can be difficult for farmers. In a country where walking is often the primary—and most reliable—means of transportation, farmers regularly walk ten miles or more to town, only to find government offices closed. 

By using UlangiziAI farmers can quickly find answers to agricultural questions, including about their livestock, plants, fertilizers and fungicides, and overall growing strategies.

“In our initial three month pilot program, we saw more than 4,000 questions come into the app in both English and Chichewa,” Essene said. “The types of questions we’re able to field, from farming practices, to crop related questions to how to test for diseases—they’re coming in a mixture modalites and the app can respond really quickly and accurately.”

To create the chatbot, Opportunity International worked with Gooey.AI, an AI workflow platform.

UlangiziAI runs on NVIDIA A100 Tensor Core GPUs in the Azure Cloud. Because Chichewa is a low-resource language, there are relatively few data assets available in Chichewa to train LLMs or speech recognition models. As a result, Gooey.AI has had to bootstrap techniques to optimize UlangiziAI. 

Video 1. A video showing the UlangiziAI app responding in real time

When a question is submitted ‌through a WhatsApp voice memo, it’s transcribed and translated into English by an LLM, such as Meta’s MMS Large or a fine-tuned version of OpenAI’s Whisper 3. When farmers upload pictures they’ve taken of a crop or pest, they are automatically routed to OpenAI’s GPT4o. That data is then combined, and sent to generate vector database queries.  

The app then queries a RAG pipeline grounded in the Malawi Ministry of Agriculture Guide to Agriculture Production Manual—a 450-page book of in-depth information focused on the specific needs of Malawian farms and farmers.

Once the retriever has identified and extracted the relevant data from the vector database, it’s sent back to the language model, which translates it into English or Chichewa. It then shares the answer back to the farmer in either text or audio format ‌through WhatsApp. Each answer also includes citations and hyperlinks to the relevant data source. 

Gooey.AI regularly hot swaps the open source LLMs and automatic speech recognition tools UlangiziAI runs on, using the near step-change improvements each successive model generation has demonstrated understanding low-resource languages. 

UlangiziAI has reduced the time it takes farmers to get critical farming guidance from days to seconds. So far, farmers who have used the app have rated it highly. And later this year or early next year, Opportunity International plans to introduce it in Kenya and make it available in Swahili as well as other Kenyan languages.

Check out Opportunity International global programming.

Learn more about the work Gooey.AI is doing to evaluate AI models and build chatbots for low-resource languages.

Discuss (0)

Tags