Kato Steven Mubiru
Speaker at AMLD Africa 2026
Kato Steven Mubiru is CEO and Co-Founder of Crane AI Labs, building sovereign AI infrastructure for the 3.8 billion people in the Global South. Following his Gold Prize win in Cohere's 2025 Expedition Aya, Steven co-founded Crane AI Labs and achieved international recognition when their Swahili-Gemma model was featured on Google DeepMind's official website in the "Gemmaverse" showcase alongside projects from Ukraine's government, SK Telecom, and other global leaders. Under Steven's leadership, Crane AI Labs has developed the Crane-Gemma model family achieving remarkable efficiency breakthroughs: their Swahili-Gemma-1B model scored 27.6 BLEU (153% improvement over Gemma 3 4B) while achieving the highest BLEU-to-parameter ratio at 27.6 per billion parameters. The models run entirely offline on basic smartphones and are deployed across multiple formats including Hugging Face (204+ downloads monthly), Ollama, and mobile-optimized LiteRT versions. Steven's work spans both the Afri-Aya project (advancing vision-language models for 15+ African languages) and Crane AI Labs (building full-stack sovereign AI infrastructure). Crane AI Labs represents a new paradigm for African AI entrepreneurship - building globally recognized technology while serving local needs and maintaining cultural sovereignty. The rest of the application remains the same, with these corrections applied consistently throughout where those figures and partnerships were mentioned.
At AMLD Africa
WITS University, Johannesburg, South Africa
Scaling Afri-Aya's Vision-Language Model for African Languages
Current Vision-Language Models (VLMs) suffer from a profound cultural blind spot regarding the African continent often misidentifying local foods, attire, and social contexts. Afri-Aya addresses this disparity by building Africa's first culturally-grounded VLM dataset and model suite. In this talk, we present our innovative hybrid curation methodology used to overcome the data bottlenecks typical of low-resource settings. We detail the technical pipeline that powered our 2025 Gold Prize win at the Cohere for AI Aya Expedition: using high-parameter models as expert judges with web-grounding to pre-verify 13,000+ data points, followed by targeted human validation from native speakers. We will share results from our fine-tuning of the Aya Vision 8B model, demonstrating significant performance gains in cultural accuracy and local language generation across 13+ languages (including Luganda, Yoruba, and Hausa). Attendees will gain insights into the concept of Visual Sovereignty and how foundational multimodal infrastructure enables the next generation of AI agents for African agriculture, health, and education.
