Abstract
<p>Abstract—Personalized online learning platforms increasinglyrely on static content sequencing and fixed assessment thresh-olds, which do not adapt instructional difficulty to a learner’sreal-time performance and engagement. This mismatch leavesfaster learners under-challenged and struggling learners over-challenged, which is associated with reduced engagement andhigher dropout in self-paced learning environments. A furthergap in the literature is that career-oriented skill development(resume-driven skill gap analysis and job-role alignment) andreal-time difficulty adaptation are treated as separate concerns,so no widely reported system combines the two. This paperpresents SkillAI, an AI-powered learning platform that formu-lates instructional difficulty adaptation as a supervised three-classclassification problem (increase, maintain, or decrease difficulty)driven by learner performance and engagement signals, andintegrates this adaptation mechanism with resume-based skill gapanalysis and personalized course generation. A Random Forestclassifier was selected for its robustness to noisy tabular features,native feature-importance interpretability, and low deploymentoverhead relative to deep or reinforcement-learning alternatives.The model was trained and evaluated on 2,000 anonymizedmodule-level interaction records logged by the SkillAI plat-form from registered learners between September 2025 and 27November 2025 (approximately three months of pilot operation),covering skill levels from beginner to advanced across multiplecourse domains. Hyperparameters were selected via randomizedsearch with five-fold stratified cross-validation, and the tunedmodel was evaluated on a held-out 30% test split (600 records)and compared against a static majority-class baseline, a rule-based heuristic baseline, Logistic Regression, a Decision Tree,and XGBoost. The tuned Random Forest achieved 95.67% testaccuracy (macro F1-score 0.9306), compared with 74.00% for thestatic baseline, 80.00% for the rule-based baseline, 90.00% forLogistic Regression, and 93.50% for a Decision Tree; an XGBoostclassifier matched the Random Forest’s accuracy (95.67%) witha marginally higher macro F1-score (0.9338), and five-foldcross-validation confirmed the result was stable (mean accuracy95.55%, standard deviation 0.53 percentage points). Quiz score,completion-time ratio, and module completion time accountedfor the majority of the model’s predictive signal. These resultsindicate that a comparatively simple, interpretable supervisedmodel, trained on real platform-usage data rather than simulatedbehavior, can support reliable, low-latency difficulty adaptation,and they provide a practical template for combining career-awarepersonalization with real-time adaptive difficulty in softwareengineering skill-learning platforms</p>