A new AI model built for the global majority
Microsoft has announced Project Gecko, a research initiative designed to close the performance gap between advanced generative AI tools and the billions of people whose languages and cultural contexts remain under-represented in today’s models. The project begins in Kenya and focuses on delivering accurate, culturally grounded and language-inclusive AI to communities that have historically been left behind by mainstream datasets.
Many generative AI systems struggle to understand low-resource languages or local cultural nuances. These gaps widen when internet content is skewed toward dominant languages and well-documented regions. Microsoft Research says this challenge explains why AI adoption remains lower in many parts of Africa and Asia, even when adjusting for connectivity and GDP.
Project Gecko aims to change this by building adaptable AI systems that speak local languages, understand local realities and run efficiently on lower-cost devices.
A global collaboration anchored in Africa
The initiative brings together experts from Microsoft Research Africa in Nairobi, Microsoft Research India, the Microsoft Research Accelerator in the United States, Digital Green and multiple partners across academia and agri-tech. The team is building cost-effective systems that incorporate speech, video and text to deliver grounded, real-world guidance at community level.
At the centre of the project is the MultiModal Critical Thinking Agent, known as MMCTAgent. This new system can analyse speech, images and video to produce answers that are context-rich and locally relevant. MMCTAgent is now accessible through Azure AI Foundry Labs, with open-source code available on GitHub for developers.
Corporate Vice President Ashley Llorens says the work reflects Microsoft’s mission to develop AI that is inclusive, responsible and globally representative. According to Llorens, building AI shaped by local languages and cultural insights leads to more useful and impactful solutions.
Why agriculture is the first frontier
Project Gecko begins with agriculture because it is a critical economic engine in Kenya and India. Millions of smallholder farmers depend on timely and accurate guidance to manage soil, pests, weather and crop cycles. However, language diversity and inconsistent digital access limit the usefulness of existing AI tools.
Farmers across both regions often switch between English, Kiswahili, Kikuyu, Kalenjin and local dialects. Many prefer oral instructions and visual demonstrations over text. Existing agricultural apps struggle because their underlying models were mostly trained on English-only datasets.
Microsoft teams built Gecko on top of Digital Green’s FarmerChat, a speech-first assistant used by millions of farmers. Digital Green holds more than 10,000 agricultural videos across 40 languages and dialects, but much of that knowledge was historically difficult to access or search.
With Project Gecko, a farmer in Nyeri County can now ask a question in Kikuyu and receive answers as text, audio or video. MMCTAgent can even jump directly to the timestamp within a community-recorded video that shows the specific solution.
How MMCTAgent delivers locally grounded answers
MMCTAgent enhances mainstream models by allowing them to reason across audio, visuals and text. It breaks a complex agricultural question into smaller parts, verifies its own interpretations and grounds its responses in real farming practices captured from community videos.
Field studies in Kenya and India show significant improvements in accuracy and usefulness. Farmers report greater trust because the answers feel locally relevant rather than generic.
One farmer from Bihar, India, shared that before using the system, advice often came from neighbours or local dealers without clear verification. With tools like FarmerChat powered by Gecko, she can follow instructions more confidently and measure results more effectively.
Building better speech tools for African languages
A major insight from Microsoft’s research is that farmers strongly prefer voice-based interactions. Yet many African languages lack reliable ASR, translation or text-to-speech tools. Project Gecko’s teams have built new speech models from scratch and have expanded support for Swahili, Kikuyu, Kalenjin, Dholuo, Maa and Somali.
The models are trained using more than 3,000 hours of Kenyan speech data and optimised for small language models. These compact systems run efficiently on low-cost devices common in rural communities, while often outperforming larger models in specialised tasks.
The team is now developing a public leaderboard to benchmark African language performance and is integrating features shaped by direct feedback from more than 130 farmers. These include clarifying questions, actionable recommendations and options that support peer-to-peer knowledge sharing.
What comes next for Project Gecko
Microsoft plans to expand Gecko’s impact across sectors such as healthcare, education and retail. However, meaningful scale will require new approaches to localised evaluation, cultural adaptation and deployment in areas with limited connectivity.
Later this year, the team will release a multilingual playbook to guide developers building domain-specific AI tools for low-resource contexts.
As Microsoft researcher Tanuja Ganu notes, the goal is to ensure that the next generation of AI is not only powerful, but also inclusive, culturally grounded and shaped by the communities it aims to serve.
The post Microsoft Launches Project Gecko to Build AI for the Global Majority, Starting in Kenya appeared first on Memeburn.
