I build and evaluate language technology and data tools for health systems in Sub-Saharan Africa.
Health guidance in Uganda is mostly written in English, and general-purpose language models can give confident answers that contradict national protocols. My work tries to close both gaps. I have built bilingual English–Luganda assistants that answer only from expert-approved or official sources, and tested them with health experts and frontline workers. I also develop open tools that help research projects map their data to shared standards.
Questions I am working on:
Grounding language-model answers in authoritative sources, such as national surveillance guidelines.
Evaluating model output with domain experts and the people who will use it.
Machine translation for health text in Luganda, including speech that mixes Luganda and English.
Predicting which machine translations need a human reviewer.
Restoring the diacritics that African-language text loses when it is typed.
Using language models to speed up mapping research data to common standards.
News
Released the Lacuna Luganda Health Corpus: 38,644 human-reviewed English–Luganda sentence pairs on health topics.
Posted a preprint on HEAL, a bilingual assistant that grounds a large language model in Uganda's disease-surveillance guidelines.
Our paper on COAST, a bilingual COVID-19 chatbot, was published in PLOS Digital Health. I am co-first author.
Co-authored a paper in Bioinformatics on building and running research computing at the Uganda Virus Research Institute.
Co-authored Ten simple rules for sustainable research computing in resource-limited settings, in PLOS Computational Biology.
Projects
HEAL: disease-surveillance guidelines in English and Luganda#
Role
Team lead; first author of the preprint
Context
ACE Uganda and the Infectious Diseases Institute, Makerere University, 2023 to present. Supported by the Gates Foundation and Code for Africa.
Uganda's surveillance guidelines are long English PDF documents, and off-the-shelf language models give plausible advice that can contradict them. HEAL retrieves the relevant guideline passages before GPT-4-turbo answers, and a fine-tuned mBART model translates between English and Luganda. Six surveillance experts rated it against three commercial models, and twenty frontline health workers gave feedback.
Infectious Diseases Institute and ACE Uganda, Makerere University, 2021 to 2023
An English–Luganda chatbot that answers COVID-19 questions only with responses approved by medical experts. It was built with Rasa, deployed on Kubernetes, and tested in two rounds with about 62 and 91 people. Intent classification reached 96.8% accuracy on 13,222 held-out queries.
Infectious Diseases Institute and ACE Uganda, Makerere University. Funded by the Lacuna Fund.
Health information in Uganda is mostly produced in English, and building translation tools for it needs vetted parallel text. The corpus holds 38,644 English–Luganda sentence pairs on health topics, each with a documented review history, and a fixed split into training, development and test sets. All 1,460 pairs flagged during review were corrected and approved, and an evaluator outside the translation team rated a sample of 1,015 pairs. The paper benchmarks two existing translation models, Sunbird and NLLB-200, on the test set.
Metadata Harmonisation Tool and the eLwazi data platform#
Role
Lead of the eLwazi Data Models, Storage, Registries and Catalogue Working Group; co-developer of the tool
Context
Uganda Virus Research Institute, 2021 to present, within the NIH-funded DS-I Africa program. The working group has had between about 20 and 50 members from DS-I Africa projects.
Research projects describe the same measurements in different ways, which makes combining their data slow. The tool uses large language models to write missing variable descriptions, suggest the matching variable in a target codebook, and give a confidence score for each suggestion.
Co-author; contributed to the ideas and edited the papers
Context
Uganda Virus Research Institute, with colleagues at ACE Uganda
How a high-performance computing service for bioinformatics was designed, run and kept going in a resource-limited setting, and the general rules other institutions can take from it.
These projects are under way. Results will be added once they are ready to share.
AfriNorm: restoring diacritics in African-language text#
Role
Lead developer; first author of the paper draft
Context
ACE Uganda, 2026. The paper draft is with colleagues for review.
Much African-language text is typed without the tone marks and other diacritics that carry its meaning. AfriNorm is one open toolkit for restoring them: a byte-level model (ByT5-small), a shared evaluation harness, and a short configuration for each language. It has been trained and evaluated on Igbo, Wolof, Yorùbá, Fon and Ghomálá', and extended to Ewe without changes to the shared code.
Translating health speech that mixes Luganda and English#
Role
Project lead
Context
ACE Uganda, started November 2026. The analysis plan was fixed before any experiments; no results yet.
In Ugandan clinics, people often switch between Luganda and English within one sentence, yet translation tools are built and tested on single-language text. This study measures how much open translation models degrade on mixed speech. It uses corpus sentences that exist both as the original mixed transcript and as an edited single-language version with the same reference translation, so the meaning stays fixed. It also tests whether Uganda-adapted models cope better than the models they were built from.
Which machine translations need a human reviewer?#
Role
Developing the study and its analysis plan
Context
ACE Uganda, started October 2026. The analysis plan is in draft; no results yet.
When our reviewers corrected machine-translated Luganda health sentences, many needed no change at all. This study tests whether open models can predict, before anyone reads a sentence, which ones need a human, so that reviewer time goes where the errors are.
Publications
Earlier publications from my MSc are listed in my CV.
Journal articles
A bi-lingual chatbot implementation for pandemic response using the transformer-based approach
Atwine MT*, Jjingo D*, Nsubuga M, et al. PLOS Digital Health, 2026. *Co-first authors.
A scalable HPC framework for bioinformatics in resource-limited settings: design principles, implementation, and sustainability from the UVRI experience
Lukyamuzi E, Kimbowa Wamala T, Ssekagiri A, Galiwango R, Kebirungi G, Atwine M, et al. Bioinformatics, 2026.
38,644 English–Luganda sentence pairs on health topics, with a review history for every pair and fixed training, development and test splits. Hugging Face, 2026. CC BY 4.0.
Mugume Twinamatsiko Atwine is a health data scientist and machine learning engineer based in Kampala, Uganda. Atwine builds and evaluates bilingual English–Luganda health assistants that answer only from expert-approved or official sources, and leads the eLwazi Data Models, Storage, Registries and Catalogue Working Group in the NIH-funded DS-I Africa program. Recent work includes a co-first-authored paper in PLOS Digital Health (2026) on a bilingual COVID-19 chatbot, and the Lacuna Luganda Health Corpus of 38,644 human-reviewed English–Luganda sentence pairs.
Funders and partners
The projects above were funded by, or carried out with, these organisations.
