CondensaCore.
A YouTube summarizer that fetches captions via youtube-transcript-api and condenses them with GPT-4 into accurate, skimmable notes — turning hour-long lectures into a few minutes of reading. Built before Gemini shipped its own summarizer feature.
Built at ConUHacks IX, a 24-hour hackathon at Concordia in January 2024. The pitch: paste a YouTube link and get a clean, skimmable summary of the lecture — a few minutes of reading instead of an hour of audio.
Under the hood it pulls captions via youtube-transcript-api, splits them into chunks the model can hold, and asks GPT-4 to compress each chunk while preserving the through-line. The frontend is a single React page that streams the result back as it lands.
This was before Gemini bundled the same feature into YouTube. At the time, the gap between "I have an hour of audio" and "I have a one-page summary" was real, and the first thing I'd built end-to-end with an LLM in the loop. The lesson that stuck: prompt design and chunking strategy matter more than model choice.