Tunga Agricultural Chatbot

Most of Rwanda’s smallholder farmers don’t have reliable internet access, but nearly all of them have a phone. Tunga is a telephone voicebot that meets them where they are: farmers call in, ask a question in Kinyarwanda, and get an answer grounded in verified agricultural knowledge - no app, no data plan, no English and no IT skills required. The system now supports the call center of Rwanda’s Ministry of Agriculture, helping staff and farmers alike get faster, more consistent answers.

The Challenge

Voice AI for a low-resource language like Kinyarwanda is a different problem than building a chatbot for English or German. Off-the-shelf speech and language tooling barely exists for it, phone-quality audio adds noise on top of that, and the answers still have to be accurate - bad agricultural advice has real consequences for a farmer’s livelihood. On top of the technical challenge, the system had to work inside an actual government call center, with real staff, real call volumes, and no room for a fragile prototype.

My Approach

I advised the project end to end. I specified the underlying technical architecture - a retrieval-augmented generation (RAG) pipeline paired with the telephone interface - and solved practical problems along the way, including keeping conversation context across turns in the chat. One piece I’m especially proud of: together with my partners, I co-designed a rigorous evaluation protocol that demonstrated the chatbot was ready for production and helped build trust with stakeholders. I also contributed across the wider delivery work - operational concept, training, documentation, and handover.

Outcome

Tunga is now live and answering farmers’ questions as part of the Ministry of Agriculture’s call center operations, in their own language, over a plain phone call. Beyond the immediate service, the project became a proof point for what’s possible in Kinyarwanda voice AI, and it directly informed later work I led on building foundational NLP infrastructure - data, models, and evaluation tools - for African languages more broadly.

Open-Source Releases

As part of the Tunga project - with Rwanda Centre for the Fourth Industrial Revolution (C4IR), KiNLP, Digital Transformation Center Rwanda, and the Ministry of Agriculture & Animal Resources - we released a suite of free open-source AI models, datasets, and source code:

• Text-to-speech model - currently the highest-quality synthetic voice for Kinyarwanda
• KinyaCOLBERT Free - the first Kinyarwanda embedding / information-retrieval model, essential for building RAG chatbots such as Tunga
• Voice dataset - for training TTS and speech-to-text (STT) models
• Information retrieval dataset - the data used to train the KinyaColBERT model
• KinyaBERT - a foundational BERT-style model serving as the backbone for various Kinyarwanda NLP tasks
• Source code - all relevant code for the models and datasets above

Technologies: RAG, IVR telephony integration, LLMs, Kinyarwanda NLP, Python.