The project
Rewatching two hours of recording to find five minutes of explanation is wasted time. This pipeline reduces each lecture to navigable written material, in five stages that can run one at a time or in sequence.
Audio is extracted from the video and transcribed with Whisper. That is the only step that runs elsewhere, on Colab or Kaggle, for the GPU: everything else stays local. The raw transcript is then cleaned by a model that reads the course context from a configuration file — title, instructor, audience, glossary, objectives — so technical terms stay correct instead of being mangled phonetically.
Two different processes run over the cleaned text: one produces multi-section notes with examples and study questions, the other typesets an annotated LaTeX handout following a fixed specification that marks filler words, questions to the class and the instructor's corrections, compiling it to PDF when pdflatex is available.
Models are chosen per stage: the lighter one for cleaning and summarising, the more capable one for typesetting, which is the task that degrades worst under simplification. Directories and models are configured through environment variables, without touching the code.
A single runner executes the local stages in order and stops at transcription: once the Whisper files are in place, rerunning it picks up where it left off.
