Design and Evaluation of an Inclusive AI-Driven Learning System for Personalized Technical Skill Acquisition in Resource-Constrained Environments
Abstract
Artificial Intelligence (AI) presents transformative opportunities for educational personalization, yet traditional adaptive learning platforms remain constrained by high bandwidth requirements, complex cloud dependencies, and standardized linear instruction. In resource-constrained technical and vocational education and training (TVET) settings, these dependencies exacerbate inequality. This study presents the design, architectural prototype, and preliminary in-silico evaluation of a lightweight AI-driven learning system tailored for dynamic technical skill acquisition. Following Design Science Research (DSR) methodology, the system utilizes a Random Forest classifier to profile learners into four competency tiers (Beginner, Intermediate, Advanced, and Expert) based on multidimensional engagement and performance indicators. Benchmarking on a synthetic dataset (N = 14,003 learner records) demonstrates that the pre-examination diagnostic model—relying strictly on continuous behavioural engagement metrics (e.g., assignment completion, attendance, study hours) without exam scores—achieves 92.4% classification accuracy (F1-score = 0.924). Incorporating post-hoc summative examination scores yields a 100.0% classification result, confirming deterministic mathematical reconstruction of established grading rules. Feature importance analysis shows low predictive weights for Gender (3.35%), Internet Access (0.78%), and EduTech Access (2.97%), though Age accounts for 11.89% of predictive weight in pre-examination profiling. Quantitative benchmark measurements confirm a lightweight model footprint (< 15MB) and minimal inference latency (< 12ms), supporting potential offline edge deployment. These findings demonstrate the architectural feasibility of early pedagogical intervention in low-bandwidth settings, while highlighting the necessity of live pilot trials for operational validation.