Background: Challenge-Based Learning (CBL) courses generate a wealth of unstructured student artifacts (e.g., reports, prototypes) that contain rich evidence of professional competencies (CDIO Standard 11). However, extracting these process variables for longitudinal analysis has historically required prohibitive manual effort, leaving valuable curriculum insights inaccessible. As a result, assessment practices often rely on coarse indicators (e.g., grades), limiting evidence-based feedback on innovation and professional skills. Challenge: While Generative AI offers a scalable solution to automate this extraction, institutions face a dilemma: utilizing frontier cloud models risks violating data sovereignty regulations, while privacy-preserving local models have historically lacked the reasoning capability required for complex assessment. Methodology: This paper presents a Privacy-Adaptive Dual-Pipeline framework to resolve this tension. We analyze a subset of 15-year archive of engineering capstone projects (N = 63 validated samples) by benchmarking a local, offline Large Language Model (LLM) against a frontier Cloud model. Results: We identify a distinct Inference Gap: while local models successfully extract explicit identity metadata (Title: 78.2% match), they struggle with semantic reasoning tasks (e.g., Value Proposition: 0.0% match) compared to the cloud benchmark. Contribution: We propose a tiered implementation strategy that allows educators to select the appropriate privacy tier based on data sensitivity. This methodology provides the CDIO community with a validated workflow to transform unstructured repositories into structured datasets, enabling scalable, evidence-based assessment and feedback without compromising student privacy.
