I build machine learning systems, mostly around feature selection, and mostly aimed at problems in Nigeria and the wider region: maternal health triage, low-resource language NLP, weather prediction for agriculture.
A lot of my work over the past year has been metaheuristic feature selection. Firefly, PSO, mutual information, applied to SVMs, random forests and logistic regression, and measured honestly. The interesting result is usually not a higher accuracy number, it is a much smaller feature set at the same accuracy, which is what actually matters when the model has to run somewhere cheap.
| Project | What it is |
|---|---|
| fake-news-firefly-svm | Binary Firefly feature selection on TF-IDF, 94.83% accuracy on 28.1% fewer features |
| weather-mi-vs-pso-random-forest | Filter vs wrapper feature selection, head to head, with the timing cost made explicit |
| MAMASAFE-AI | Maternal health risk triage, tuned for recall on the high risk class rather than accuracy |
| stroke-prediction-firefly-svm | SVM hyperparameters tuned by a firefly swarm instead of grid search |
| PROJECT180 | 180 days from data engineering to AI engineering, logged daily in public |
Working through Project180, a 180 day run into AI engineering, with local data infrastructure for maternal health and low-resource NLP as the through-line. I also mentor final year students through their machine learning projects.