T-ETHER

Upload a video -> get chapters, key moments, and short-form reels
A library for all uploaded videos
AI generated hook to a reel: rewords the key topic covered into a question and put into TTS
Creates a sub ~90 second reel that explains one of the key topics detected by clipping only related video segments into a seamless video

Our attention spans are getting destroyed by the internet. Short-form feeds are engineered to hijack focus, and even when people want to learn, study, or build something meaningful, they end up trapped in an endless scroll optimized for distraction, not depth.

We kept asking ourselves: What if we could redirect doomscrolling into something productive, without the brain rot?

ENTER T-ETHER.

T-ETHER transforms long educational videos into binge-worthy micro-lectures. Upload a 50-minute lecture, get back a series of 60-second TikTok/reel-style clips, each one focused on a single concept, complete with hooks, animated captions, and smart visual editing. No more scrubbing through hour-long recordings to find the content that actually matters. T-ETHER automatically identifies key topics, extracts the most relevant moments, and packages them into short-form content that's actually easy to watch and retain. For students: Turn any lecture into a playlist of digestible clips you can watch between classes, revisit before exams, or finally get through without zoning out. For creators: Repurpose long-form content into scroll-stopping clips without hours of manual editing. The original message stays intact, we just make it impossible to ignore.

HOW IT WORKS.

T-ETHER runs on an end-to-end pipeline from upload to polished export:

Upload & Ingest: The user drops in a video through our React frontend; the backend queues the job and kicks off processing.

AI Analysis: We use TwelveLabs Pegasus and Marengo in conjunction: Pegasus identifies the important key concepts in the video, and creates semantic search prompts for Marengo to shift through the entire video and search for the most relevant moments for each topic.

Smart Clipping: After Marengo gives us our timestamped segments, we smoothly concatenate the relevant clips by specifically extracting segments that align with natural speech patterns using Deepgram’s speech-to-text to avoid cutting off professors mid-sentence.

Audio & Captions: Deepgram generates transcriptions and text-to-speech for hooks. Animated subtitles sync word-by-word for that native short-form feel.

Render & Deliver: FFmpeg stitches everything together with smooth fade transitions, normalized audio levels, and a clean 16:9 export ready to share.

THE CHALLENGES:

Manipulating the APIs to “work how we wanted it” was definitely not easy, especially from intense hackathon pressure. We had to learn how to be specific in our utilization of Marengo and Pegasus so they would compliment each other flawlessly. We ensured to create specific queries from Pegasus's generative model to ask the Marengo foundation model to clip the important relevant segments via its semantic encoding and video search capabilities.

WHAT WE LEARNED:

Attention isn't a willpower problem, it's a design problem.

We started building a video tool. We ended up learning about human psychology. The internet has trained our brains to expect dopamine in the first 2 seconds. If your opening frame feels like work, users bounce, even if what comes next could change how they learn forever. The insight: don't fight the scroll, redirect it. Make the hook irresistible, make the content effortless, and suddenly "educational" stops feeling like a chore. We also realized the best hackathon projects aren't feature-rich, they're behavior-accurate. Every design decision should answer one question: does this make the right action the easiest action?

Don't build tools that demand attention. Build experiences that capture it.

WHAT’S NEXT:

Personalized learning paths: Users choose topics they want to focus on, skip what they already know, and get micro-lectures tailored to their goals

Smarter visuals: Pull in relevant images, diagrams, and clips from across YouTube and the web, not just the source video

Multi-video synthesis: Combine concepts from multiple lectures into unified study playlists

Beyond lectures: Expand to podcasts, documentaries, tutorials, and any long-form content worth learning from.

Our goal stays the same: make education as addictive as doomscrolling.