01
Low-resource language NLP
Natural language processing for African languages with little or no representation in existing corpora, models and benchmarks — including expanding the language coverage of large language models.
Research
Language technology for the languages that don't have any
I am a graduate researcher in the Department of Computer Science at Makerere University, working on speech and translation for Ugandan languages — and on the question of whether those systems are safe enough to put in front of people.
Current work
Speech-to-speech translation for Luganda
My main project is an end-to-end speech-to-speech translation system for Luganda. The interesting part is not the pipeline — it is that almost every assumption modern speech translation rests on fails here. There is no large paired corpus. There is no established benchmark. The models that work well elsewhere were trained on data that does not include this language at all.
A translation system that is 90% right is useful for a tourist and dangerous in a clinic. Deciding which of those you are building is a research question, not an engineering one.
So the work runs in two directions at once. One is adaptation: taking architectures like Wav2Vec and Whisper and making them work under conditions they were never designed for. The other is evaluation and safety — building the frameworks that tell you honestly how good a system is, and detecting the ways it fails.
That second direction matters more than it usually gets credit for. I am investigating bias detection in translation systems, evaluation for high-stakes contexts such as medical and legal communication, and alignment in cross-cultural settings. I work with linguists and domain experts throughout, because a system can be technically correct and still culturally wrong in ways no metric will catch.
Underneath all of it is a fairly simple commitment: democratising access to AI for East African languages, so that the people who speak them are not permanently downstream of technology built for someone else.
Research interests
What I work on
These overlap more than the headings suggest — most of my work sits in the intersection of two or three of them.
02
Speech recognition & translation
End-to-end speech-to-speech translation, automatic speech recognition, and the practical question of how much speech data these systems actually need before they become useful.
03
Safety, bias & evaluation
Bias detection in translation systems, evaluation frameworks for high-stakes contexts such as medical and legal communication, and alignment in cross-cultural settings.
04
Multilingual model development
Adapting transformer architectures to low-resource conditions, training over large-scale datasets, and the distributed training techniques that make that tractable.
05
Production ML & MLOps
Getting research systems into production and keeping them there — pipelines, deployment, versioning and monitoring, so that a result survives contact with real usage.
06
Explainable computer vision
Interpretable models for agricultural disease detection, where a prediction is only actionable if the person receiving it can understand what it is based on.
Applied research
Where the research has gone
Research that stayed on paper is only half finished. These are the places the work has actually been deployed or published.
Sunbird AI
Translation & speech in production
Sunbird's translation and speech recognition models run in production serving
real-time inference, with data pipelines, monitoring, and BLEU/chrF/COMET
evaluation built around them.
See the case study
Makerere AI Lab
Crop disease detection
Computer vision for Black Sigatoka in banana plantations, published twice with
Springer in 2023 and shipped inside a mobile application for use in the field.
See the publications
Emergent AI
Computer vision for health
Machine learning and computer vision applied to health-sector datasets, with feature extraction pipelines for large-scale medical imaging and validation workflows to support data-informed health insights.