Developed a Python / scikit-learn pipeline that pits PCA against LDA on the UCI HAR dataset (10k traces, 561 features). Automated preprocessing, 5-fold cross-validation, and grid search showed that a 5-dimensional LDA with an SVM-RBF cuts features by 98% while scoring macro-F1 0.984 — a +16 pp gain over the raw baseline.
The result: resource-efficient activity recognition suitable for wearable and mobile deployments where memory and latency budgets are tight.