AnITa AI Assistant for Georgia’s digital education platform

GITA – Georgia’s Innovation and Technology Agency – is responsible for advancing technology adoption and digital inclusion across the country. Working under the EU-funded AnITa programme, GITA commissioned the development of an AI-powered assistant for Georgia’s national learning platform. Civitta won the engagement through a public tender and delivered it in partnership with GeoLab, the team behind the platform itself.

The brief was clear: build something accurate and genuinely useful for students, with limited room for iteration.

• The core technical challenge was the Georgian language. No large language model handles Georgian reliably out of the box, making an off-the-shelf solution unviable.
• The engagement also carried a fixed budget, together with the deadline, and required two independent organisations to deliver as one. 

Civitta approached the engagement with a clear principle: build only what matters, and build it to last.

Rather than attempting to train a model on Georgian from scratch, the team designed a smarter architecture. The assistant operates in Georgian at the surface level, using translation and a leading language model in the background – delivering fluent, accurate responses without the cost of a custom-trained model.

The core design decision was how the assistant determines what to answer. Civitta built a three-layer retrieval system:

• first checking a curated set of trusted answers to the most common student questions;
• then searching the actual curriculum materials;
• and only falling back on the model’s broader knowledge if neither layer produces a match.

This approach keeps responses grounded in real course content and minimises the risk of inaccurate or irrelevant answers. The system was also designed for reuse – built as a product from the outset, not a one-off solution tied to a single platform.

Delivery required close coordination between Civitta, GITA and GeoLab. With two independent teams working as one, consistent communication and proactive relationship management were central to navigating the project – including adjustments required by funding reductions – without compromising the deadline.

The assistant was delivered on time and integrated into Georgia’s national learning platform, providing real-time, curriculum-grounded support to thousands of students with capacity to handle hundreds of concurrent users.

Beyond the immediate delivery, the three-layer architecture was designed as a reusable product from day one. It has since been adapted for other platforms and sectors, demonstrating that the constraints of the original brief produced a more focused and transferable solution than an unconstrained build would have.

Services provided:
AI and Data, Software development. Built with FastAPI and PostgreSQL, Chroma and Azure AI Search for retrieval, Celery and Redis for async processing, OpenAI with Azure Translation for the multilingual layer, and Docker for deployment.

1. Constraints produce sharper products: A fixed budget and firm deadline forced focus on what mattered most – resulting in a more robust and reusable system than an open-ended brief might have delivered.

2. Small, aligned teams outperform large ones: Clear shared purpose and consistent communication between Civitta, GITA and GeoLab drove delivery more effectively than formal process structures.

3. Design for reuse from day one: The three-layer retrieval architecture was conceived as a product, not a one-off. Any organisation dealing with high volumes of repetitive questions or needing to make information accessible at scale is a natural fit.