There is a lack of professional rehabilitation therapists and facilities in low-resource settings such as Bangladesh. In particular, the restrictively high costs of rehabilitative therapy have prompted a search for alternatives to traditional in-patient/out-patient hospital rehabilitation moving therapy outside healthcare settings. Considering the potential for home-based rehabilitation, we implemented a low-cost wearable system for 5 basic exercises namely, hand raised, wrist flexion, wrist extension, wrist pronation, and wrist supination, of upper limb (UL) rehabilitation through the incorporation of physiotherapists' perspectives. As a proof of concept, we collected data through our system from 10 Bangladeshi participants: 9 researchers and 1 undergoing physical therapy. Leveraging the system's sensed data, we developed a diverse set of machine learning models. and selected important features through three feature selection approaches: filter, wrapper, and embedded. We find that the Multilayer Perceptron classification model, which was developed by the embedded method Random Forest selected features, can identify the five exercises with a ROC-AUC score of 98.2% and sensitivity of 98%. Our system has the potential for providing realtime insights regarding the precision of the exercises which can facilitate homebased UL rehabilitation in resource-constrained settings.