
SketchTeach reimagines digital education by providing a personalized, "professor-at-the-whiteboard" experience for any subject. The project addresses the gap between static LLM text responses and engaging visual learning. When a user asks a question, SketchTeach triggers a sophisticated multi-agent workflow coordinated via FastAPI. First, an Explainer Agent (Qwen2.5:14b) breaks the topic into a structured 4-step lesson. Simultaneously, a Voice Agent generates natural narration using edge-tts. These components feed into an Animator Agent, which writes a custom Manim Python script to render concept-specific diagrams—like binary trees for DFS or client-server boxes for TCP—perfectly synced to the audio duration. The entire backend is powered by AMD Instinct MI300X GPUs via the AMD Developer Cloud. By leveraging ROCm and the massive HBM3 memory of the MI300X, we achieve rapid inference for the 14b model and high-speed parallel rendering. This allows the system to move from a raw query to a fully produced, high-fidelity instructional video in under 60 seconds. SketchTeach isn't just a RAG wrapper; it’s an autonomous content engine that handles logic, creative visualization, and self-correction to provide a seamless educational experience.
10 May 2026