CobbleAI: course-grounded adaptive learning with local LLM inference, retrieval-augmented generation and automatically constructed knowledge graphs

Introduction: Learning management systems store course content as static archives, and widely deployed course-tutoring systems route student data through third-party cloud services, excluding institutions with data-residency obligations, unreliable connectivity or no budget for per-token charges. We ask whether privacy-preserving, course-grounded intelligent tutoring can be delivered on low-cost, institution-controlled infrastructure. Materials and methods: CobbleAI runs retrieval, knowledge graph construction, tutoring inference, assessment and behavioural analytics entirely on institution-owned hardware. It combines a four-stage retrieval pipeline (query expansion, batched bi-encoder retrieval, deduplication and cross-encoder reranking with weighted score fusion), two-pass multi-document knowledge graph construction, a Gemma 4 E2B model (2.3 billion effective parameters) adapted with Low-Rank Adaptation on a single consumer graphics processing unit (GPU), an interaction-derived confusion score and a graph-constrained study planner. A four-week feasibility pilot involved 48 students in three computer science courses. Results: Retrieval-path latencies were 340 ± 35 ms for bi-encoder retrieval, 180 ± 22 ms for reranking and 530 ± 48 ms end to end; graph construction took 88 ± 12 s for a fifteen-document course. Query expansion and time-to-first-token were not instrumented, so user-perceived conversational latency is not established. A provisional retrieval benchmark returned precision@5 of 0.54. The mean System Usability Scale score was 67.92 (standard deviation (SD) = 14.01), not different from the benchmark of 68, with 43.75% of participants in the marginal or poor bands. Instructors confirmed seven of nine concepts flagged as high-confusion (78%; 95% confidence interval (CI) 45.4–93.8%). Conclusions: A locally hosted model of about two billion effective parameters can sustain the evaluated course-grounded tasks on commodity hardware without external inference, with retrieval and reranking at interactive latency. The pilot establishes technical feasibility and adequate usability, not educational efficacy; a randomised controlled evaluation is required to assess learning outcomes.