An AI customer support centre that speaks Kinyarwanda

On 12 June 2026 I defended my capstone project: an AI-powered customer support centre that works over the web and over VoIP phone calls, in both English and Kinyarwanda.

My academic supervisor was NIYONSABA Vedaste, with guidance from NIYONSHUTI Yves.

Why Kinyarwanda matters

Most AI tools work best in English. In Rwanda, many customers prefer, or only speak, Kinyarwanda. A support system that only works well in English leaves out the people a business most needs to reach.

Two channels

The system answers customers in two ways:

• On the web, through a chat interface.
• Over the phone, through VoIP, so a customer can call and speak to the AI.

The phone channel matters for the same reason USSD mattered for Uruziga: not every customer will use a website, but almost everyone can make a call.

How a call is answered

The pipeline is simple to describe:

1. The system receives the caller's voice.
2. It transcribes the speech into text.
3. It uses retrieval-augmented generation (RAG) to find the relevant information in the business's knowledge base and compose an answer.
4. It converts the answer back into speech with text-to-speech, so the caller hears it.

The hard part

The most complex part was making VoIP and large language models work together. A phone call happens in real time. Transcription, retrieval, generation and speech all have to fit inside the pause a caller will tolerate, over a voice connection that doesn't wait.

My earlier work helped. In January 2026 I had put custom VoIP and two-way SMPP software into production for Trust SMS Uganda, so the voice side was familiar ground.

What I took from it

Language is access. Supporting Kinyarwanda is not a feature; it decides who can use the system at all.

Voice is still the most universal interface. Where a call route is set up, a phone call needs no app, no data plan and no English.

The model is only part of the system. Getting calls in and spoken answers out, fast enough to feel like a conversation, is where most of the engineering goes.