05
Jarre
youtube-to-course generation, voice tutoring, and rubric-based evaluation
Nicolás built Jarre as an open learning platform that turns YouTube videos into structured courses with generated lessons, quizzes, timestamps, voice tutoring, and spaced repetition.
- 44+ Supabase tables
- 6 pipeline stages
- 7 voice modes
Architecture
The interesting part of Jarre is not the wrapper. It is a staged learning product with explicit pipeline limits, persisted state, evaluation layers, user-knowledge sync, and a realtime voice runtime instead of one loose tutoring prompt.
Evaluation system
Jarre tries to evaluate understanding instead of memorization. It uses 10 specialized rubrics with three dimensions each, a mastery ladder, and concept-level memory for strengths and misconceptions rather than relying on one generic LLM score.
Voice system
The voice layer is a productized realtime transport. Jarre sends PCM 16 kHz audio into Gemini Live, receives 24 kHz audio back, handles server-side VAD, tool calls, reconnection, transcript callbacks, and seven session modes so spoken tutoring stays structured.
Pipeline runtime
Jarre treats ingestion as a runtime contract. It accepts YouTube and direct PDF sources, builds deterministic user-scoped resource ids, skips reruns when a resource is already processed enough, checks token budget before starting, and mirrors finished resources back into user knowledge.