Public institutions want to use generative AI for citizen services but often lack the budget, technical capacity, and infrastructure to do so without getting locked into proprietary vendors. To help close that gap, ITU built GENIE.AI, a low-cost, open-source software framework designed specifically for government use, and assembled the system using existing open-source components. They used OPEA to manage the workflow that connects the AI to specific government documents, a process known as retrieval-augmented generation. They integrated Docling to extract text from complex PDFs, which are common in government records. For the actual text generation, they used IBM Granite, specifically the granite-4.1-8b model, and Google’s Gemma for translation and safety checks. They optimised the framework to run these smaller models efficiently so the system can operate in environments with limited computing resources. A global competition then used the framework to build public-sector applications.
This open-source stack is powering pilot deployments for real-world government services in Lesotho, the Gambia, and Bangladesh. Out of more than 300 submissions from 79 countries, teams produced three winning pilot programmes now entering a three-month deployment phase. These include an agricultural assistant in Lesotho, a health guide in The Gambia, and a weather advisor in Bangladesh. A separate pilot is also running in El Salvador to help agriculture extension workers. The authors note that custom enhancements to the workflow help compensate for the performance gap between smaller models and larger systems. Open-source AI technologies become far more valuable when they are assembled into reusable public digital capabilities that governments can own, adapt, and scale.
Today’s other reading is in assorted links.