Gilbert Yiga

I am a machine learning engineer working on natural language processing for low-resource African languages. I split my time between Sunbird AI, where I design and run production ML systems, and Makerere University, where I am a graduate researcher in the Department of Computer Science.

Most language technology is built for languages that already have plenty of data. I am interested in what happens when they are not — and in making sure the systems we build for those languages are good enough to trust.

My research focuses on speech-to-speech translation for Luganda, and on the safety questions that come with it: bias detection, evaluation frameworks for high-stakes contexts like medical and legal communication, and alignment across cultures. In practice that means adapting architectures such as Wav2Vec and Whisper to settings they were never trained for, and building the evaluation harnesses — BLEU, chrF, COMET — that tell you honestly whether it worked.

The engineering half matters just as much. A model that only exists in a notebook helps nobody, so I build the data pipelines, serving infrastructure and monitoring that take a model from a research result to something running reliably for real users. I also work closely with linguists and domain experts, because the hardest problems in this work are rarely only technical.

Before this I built enterprise CRM systems at Pahappa, applied computer vision to health datasets at Emergent AI, and worked on crop-disease detection at the Makerere AI Lab — which became my first two publications.