Building Kikuyu AI: Interview with C-elo's Research Team
January 08, 2026
Research Team Interview
Founder & Lead Researcher
Mark Gatere
How do you build a chatbot that speaks Kikuyu as fluently as ChatGPT speaks English? Our research team discusses the technical challenges of fine-tuning LLMs for low-resource Bantu languages and our roadmap for voice AI.
overview
C-elo Labs was founded with a singular mission: to build AI that speaks African languages natively. We started with Kikuyu because it presents fascinating technical challenges—tonal semantics, agglutinative morphology, and code-switching with English and Swahili.
Our approach involves fine-tuning Google's TranslateGemma 12B using QLoRA and DoRA, transforming 30K translation pairs into instruction datasets, and generating 50K+ synthetic conversational examples. We're also building Speech-to-Speech models using the Mimi neural codec for real-time voice interactions.
This interview explores the technical decisions behind our work: why TranslateGemma over alternatives, how we handle typos and missing diacritics, and our roadmap for expanding to Kamba, Luo, and other Kenyan languages.
Interviewees:
Mark
Gatere
Founder & Lead Researcher
AI researcher focused on low-resource language NLP. Leading C-elo's efforts to bring high-fidelity AI to African languages, starting with Kikuyu.
——Why did you choose Kikuyu as the first language?
Gatere: Kikuyu has over 6 million speakers and a rich oral tradition, but almost no AI support. It's also my mother tongue, so I can validate quality directly. The language has complex tonal patterns and agglutinative morphology that make it an excellent test case for our techniques.
——What's the biggest technical challenge you've faced?
Gatere: Tokenization. Standard tokenizers fragment Kikuyu words into meaningless pieces because they were trained on English. A word like 'tũtikũmũhe' (we will not give him) gets split into 7+ tokens. We chose TranslateGemma specifically because its 256K vocabulary handles Bantu morphology much better.
——How do you handle typos and missing diacritics?
Gatere: We inject noise into our training data. For every clean example, we create noisy variants—removing diacritics (ĩ→i), simulating keyboard errors, dropping vowels. This teaches the model to map messy input to correct output, just like English models handle typos gracefully.
——What's next for C-elo Labs?
Gatere: Voice. We're building a Speech-to-Speech model using the Mimi neural codec so farmers can ask questions in spoken Kikuyu and hear answers instantly. We're also preparing to expand to Kamba and Luo. The goal is an AI that truly serves Kenyan and African communities.
