Open-Source AI

Reimagining AI for low-power devices among under-resourced communities

Our projects are dedicated to advancing inclusive AI through the revitalization and preservation of low-resource languages in Africa. We are currently developing an AI-enhanced SMS system to support the Ogiek community’s efforts to protect and transmit their language and knowledge systems. The Voice of the Ogiek / Sauti Ya Ogiek / Salautab Ogiek initiative reflects our broader commitment to non-extractive, participatory, and community-driven research.

We are currently seeking funding to expand this work to additional low-resource languages in Kenya. Our hope is that communities across Africa will be able to adapt and implement the open-source code, guides, checklists, and toolkits we’ve developed.

All of our projects are grounded in a non-extractive, participatory, and community-led framework. Our projects are co-designed with the communities, run on Low-power AI, and are on local servers and not in the cloud, for (a) data sovereignty of indigenous communities, (b) ensuring trust, safety, and privacy, and (c) climate justice as low-power AI runs on low compute. 

The aim is to advance free and open-source, community-engaged technology initiatives by focusing on FOSS, low-power computing, data justice, and locally deployable AI systems. Our goal is to collaborate with under-resourced communities to co-design accessible and culturally grounded AI tools that align with local priorities. Rooted in the principles of decoloniality and epistemic justice, particularly regarding African Indigenous knowledge systems, the project interrogates how AI can be developed not to replicate existing global power hierarchies but to empower communities in shaping their digital futures.

One of our primary objectives is to develop representative multimodal datasets, encompassing text, image, audio, and video, to train open-source large language models (LLMs) that accurately reflect the lived realities, languages, and values of diverse African contexts. These datasets are essential for building AI systems that are both inclusive and responsive to the cultural and linguistic diversity of historically marginalized communities.

Key initiatives include a project by members of the collective titled “Wezesha na Kabambe: Swahili Voice Chatbot for Empowering Smallholder Farmers in Kenya,” which was funded by Mozilla Common Voice and focused on co-creating a proof-of-concept for a mobile-enabled Swahili chatbot that functions on both feature phones and smartphones. Powered by a locally hosted database of frequently asked questions, the chatbot provides critical information access in environments with limited internet or electricity. Another ongoing project by some members of the collective is titled AI-Generated (AIG) Short Message Texts (SMS) for Environmental Justice among the Ogiek Indigenous Community in Mau Forest, funded by the Mozilla Technology Fund. This project involves co-designing an AI-powered SMS prototype that generates SMS messages for land rights and environmental justice. Both of the above initiatives are engineered to function effectively in low-resource settings, prioritizing accessibility and offline usability.

To support these initiatives, we are exploring Local AI approaches (a form of edge AI) that enable large language models (LLMs) to run on low-power devices such as CPUs and feature phones, without internet connectivity or cloud infrastructure. By leveraging tools like Ollama, an open-source platform that allows LLMs to be executed locally, we ensure greater data privacy, security, and sustainability. This approach reduces dependence on high-bandwidth internet and helps bridge the digital divide in remote or infrastructurally underserved areas. One such innovation is the salautabogiek prototype. A low-power AI SMS prototype for feature phones, using locally hosted LLMs optimized for low-resource languages and low-bandwidth environments.

We are also currently exploring possibilities and funding for the scaling up of the prototype’s testing by swapping out lightweight LLMs for different use cases to deploy among women farmers in Bungoma and Maasai pastoralists in Kajiado.

We also address the broader challenges facing resource-constrained communities, that is, those with limited access to healthcare, education, electricity, and digital connectivity. Traditional AI models, which often require powerful computing resources and stable internet connections, are typically inaccessible in these contexts. Our goal is to design AI systems that are lightweight, energy-efficient, and deployable on affordable hardware, thereby bringing equitable technological access to areas often overlooked by mainstream innovations.

Ultimately, our vision is to democratize the development and use of artificial intelligence. By equipping communities with the tools and knowledge to customize and govern AI systems, we foster not only digital inclusion and data justice but also community sovereignty. Our broader research contributes to the integration of African Indigenous knowledge into technology design, ensuring that AI is community-driven, adapted, and shaped by the communities it is meant to serve.