skyclass-distill

SkyClass Distill is a TypeScript-based teaching capability distillation system designed to compile real classroom videos, verbatim transcripts, and keyframes into traceable, structured Teaching Skills. The system implements a complete pipeline for converting raw media into teaching assets and provides a restricted Pi Agent runtime to execute these skills and deliver instructional explanations. Core capabilities include: 1. Media Processing: Integrates yt-dlp, FFmpeg, and whisper.cpp to support public video collection, audio extraction, keyframe capture, and remote or local audio transcription. 2. Knowledge Distillation: Utilizes LLMs to extract structured information—such as teaching actions, expected responses, and learning checks—from classroom evidence to generate Skill packages containing source evidence. 3. Runtime and Evaluation: Provides a restricted Pi AgentSession runtime that opens only specific tools to ensure the teaching process is auditable; includes a built-in evaluation center that supports comparing the instructional effectiveness of different skill configurations (e.g., plain text vs. visually enhanced) via versioned benchmarks. 4. Technical Architecture: Employs a React 19 + Vite frontend and a Fastify Node Server backend, using atomic JSON files for storage to ensure data-code separation and an auditable chain of custody.