A Multimodal Conversational Agent for Co-Creating Teaching Slides and Lesson Plans

Abstract

This paper presents a multimodal conversational teaching agent that co-creates classroom slide decks and lesson plans with teachers through a staged review process. The prototype was motivated by a practical teaching-design scenario in which courseware preparation is fragmented across several tools and teacher intent is often captured only superficially. The proposed system combines three capabilities: multi-turn intent clarification, retrieval-augmented generation over a locally curated knowledge base, and optional reference-file grounding from uploaded PDF, DOCX, PPT, image, or video materials. Instead of exporting final artifacts at the first interaction, the agent follows a text-first pipeline. In Round 1 it produces a slide draft and a lesson-plan draft for review. In Round 2 it converts the approved slide draft into a structured slide object and then into PPTX. In Round 3 it converts the approved lesson-plan draft into Markdown and then into DOCX. This design keeps the teacher in control of pedagogical structure, terminology, and difficulty level while preserving the efficiency gains of large language models. The implementation separates chat-oriented intent handling from deterministic formatting and export steps, and it uses explicit state variables to support revision, regeneration, and artifact continuity across rounds. A case-based verification on networking topics such as VLAN, WLAN, and STP shows that the system can align courseware drafts, exported slides, and exported lesson plans under a single interaction loop. The paper contributes a reproducible low-code architecture, a multimodal fusion strategy that does not require writing uploaded references into the knowledge base, and a teacher-centered workflow for controllable educational content generation.