Namu is an African AI research and technology company building speech-native models, datasets, and products that make technology accessible through natural conversation in every language and every community.

Assisted a PhD student in applying pruning and compression techniques to transformer-based LLMs using Python and PyTorch, evaluating their impact on model size, memory usage, and inference speed. Supported the experimental workflow by preparing datasets, running training and pruning scripts, and monitoring GPU performance across different sparsity levels.

I have built a centralized plaform for the UwcNiger community where current students can find guidance, connect with alumnis. The platofrm offers a sense of community that will help us stay connected and help each other.

Working with 1,000+ students globally on collaborative research initiatives, under the mentorship of faculty and researchers affiliated with MIT, Georgia Tech, UGA, Emory, and other institutions, contributing to research papers aimed for global publication.

Built a fine-tuning pipeline for Hausa text-to-speech by preparing a single-speaker dataset and adapting Meta's MMS Hausa checkpoint with the original VITS training modules. Reproduced MMS's character-level text processing and added targeted pronunciation overrides, working around the Transformers VITS wrapper's inference-only limitation.

I won the 1st place ($3k + ticket to the award ceremony) on the Financial Stress Solution Hackathon on Zindi. My solution predicts the probability that a mobile money customer will experience liquidity stress in the next 30 days. It uses the customer’s last 6 months of behavior, creates features that describe changes in income, spending, balance, and activity, then trains several models and chooses the best final prediction strategy.