Learning the violin is a fine motor skill driven by a perception-action feedback loop. In weekly lessons, a teacher completes this loop by diagnosing errors and providing targeted interventions. However, during solo home practice, beginners cannot reliably detect or diagnose their own mistakes, which often leads to cementing errors rather than fixing them.

This proposal introduces an intelligent practice assistant designed to close this at-home feedback loop. While such systems can be modeled end-to-end, this work argues for a hybrid architecture that embeds explicit pedagogical structure. The central hypothesis is that under the scarce expert training data typical of music pedagogy, explicit structure acts as a powerful inductive bias. By combining learned multimodal perception with a teacher's diagnostic vocabulary of hypothesized error-to-cause attributions, a structured system can match or exceed end-to-end models in diagnostic accuracy and intervention selection while remaining highly sample-efficient and auditable. Each layer is evaluated against a matched unstructured baseline on identical data, spanning diagnosis, feedback generation, and intervention policy. An early prototype running the full perception-action loop on real recordings already supports this approach. The proposal details ongoing lesson data collection and an 18-month roadmap to evaluate the full decision policy.