This study evaluated a multimodal generative AI tutor that delivers explicit instruction on phonological processes and feature geometry to second language learners. The tutor combined language model generated explanations, speech synthesis, interactive articulatory visualizations and rule induction activities. Ninety intermediate L2 English learners participated in a mixed methods quasi experimental design across eight weeks. Participants were assigned to receive instruction from either an AI tutor or using traditional explicit instruction methods. Measures included tests of unambiguous knowledge of phonological processes such as assimilation and feature spreading, metaphonological awareness tasks as well as controlled production accuracy and transfer to free speech. Results revealed significantly greater gains in explicit knowledge and metaphonological awareness for the AI group. Production accuracy demonstrated greater improvements for process driven errors in the AI condition and greater transfer to spontaneous speech was also observed. Participants reported high levels of perceived usefulness of the visualizations and adaptive scaffolding, though a few noted occasional confabulation of rule formulations by the AI. These findings demonstrate the capacity of multimodal generative AI to bridge theoretical phonology instruction and practical production outcomes. The study provides a reusable tutor framework and evidence for the value of explicit knowledge of abstract phonological structure when delivered through adaptive multimodal instruction. Further analyses revealed that gains were particularly notable for complex phonological processes involving feature spreading across hierarchical nodes, with post-test retention remaining robust four weeks later. Exploration of interaction data also revealed that learners who engaged frequently with interactive feature tree manipulation achieved higher transfer scores, underscoring the value of active structural engagement in addition to generative explanations.