InsYouTube is practically the default university for developers and students worldwide, hosting millions of hours of world-class tutorials and lectures. Yet, the platform is fundamentally engineered for continuous entertainment, not structured retention.

During study sessions and hackathons, we found ourselves constantly scrubbing through multi-hour lectures, pausing every two minutes to jot down code snippets, and struggling to remember core takeaways the very next day. The problem wasn't a shortage of high-quality material—it was the cognitive friction of turning long-form, unstructured video into an organized curriculum. We asked ourselves: Why can't any public YouTube video behave like a structured, interactive course on Coursera or Udemy, complete with milestone lessons, synchronized notes, and checkpoint assessments? That question became the spark for our platform. piration

What it does How We Built It

We designed the architecture to handle the entire lifecycle of educational content: ingestion, parsing, pedagogical structuring, and synchronized playback.

How we built itIngestion & Transcription:

The backend takes a video or playlist URL and extracts timestamped subtitles via youtube-transcript-api. For videos lacking native captions, the pipeline extracts compressed audio tracks and transcribes them using speech-to-text models.

Challenges we ran into Challenges We Faced

Managing Long Transcripts & Token Overhead: Multi-hour crash courses can easily exceed millions of characters of raw spoken text. Passing raw transcripts naively causes severe latency and prompt truncation. We resolved this by building a time-window chunking pipeline that aggregates adjacent transcript snippets into semantic segments before passing them to the model's high-context window with deterministic schema constraints. Noisy Spoken Grammar to Technical Accuracy: Spoken tutorials contain frequent filler words, conversational detours, and speech-to-text phoneme misinterpretations (e.g., hearing "sequel" instead of "SQL" or "sink" instead of "async"). We engineered multi-stage prompt guidelines requiring the model to cross-reference technical keywords against adjacent context to guarantee syntax accuracy in code blocks. Bidirectional Player Synchronization: Keeping a custom note-taking UI seamlessly bound to the YouTube IFrame without causing stuttering or race conditions required careful state management. We debounced player polling events and decoupled active time updates from user scrubbing actions.

Accomplishments that we're proud of

What we learned What We Learned

Education Demands Structure Over Summaries: Summarization condenses text, but learning requires active synthesis. We learned that extracting bite-sized milestones and immediate checkpoint questions dramatically improves information retention compared to passive reading. The Power of Strict Schema Enforcement: Relying on open-ended LLM generations leads to brittle UI parsing errors. Enforcing strict JSON schema responses eliminated parsing exceptions and allowed the frontend to dynamically construct complex course navigation trees reliably. Engineering for User Cognitive Load: A good developer tool should disappear into the background. Synchronizing video scrubbing directly with conceptual timestamps cut down the manual effort of navigation, proving that subtle interface ergonomics often deliver the highest user impact.