Dartmouth Fall 2023 Technigala

Discite is a platform for shortform learning. We offer real-time streaming of short videos, with guiding "learning paths" and ability to engage with and share clips to friends.

Problems

Short-form content is the future of content consumption. But short-form content is not optimized for learning, at least not yet.

Instead, it’s optimized for engagement. TikTok and the likes try to lure you into doom-scrolling your time away.

We asked ourselves an important question — can a user have:

• Short-form content
• Optimized for and customized to their personal learning habits…
• Without predatory, dopamine-inducing hooks?

One may also wonder whether we should just retain long-form content for learning. A side-effect of short-form content is a shrinking attention span, which makes it increasingly tougher for learners to labor through long videos without losing interest.

• Content creators recognize this, and want to make shorter content.
• Even existing long-form content platforms such as YouTube are introducing short-form content on their platforms.

We commend that. But the current approach to serving short-form content is to randomly sequence videos on varying topics and show them to the user in a feed. These videos will jump around entirely unrelated topics, depending on what is "hot" at the moment. This sequencing is not optimized for learning, so we want to create a better solution that is created for users who want to learn and not just be entertained.

We also recognize the existence of excellent platforms such as Khan Academy, but we think there is a missed opportunity since they do not offer short-form content (sub-minute) and also have no social features.

Proposed Solution

At the start, we asked each other and some of our friends how we would envision learning in the current age of short-form content and artificial intelligence.

This journey led us to Discite. Discite allows you to:

• Stream short, atomic clips addressing specific subtopics, through a mobile app, in real-time.
• Access a "flow" of such clips through a learning-path. No need to manually identify the next subtopic and the relevant clip.
• Engage with such clips and learning-paths.
• Add friends and share insightful clips or learning-paths with them.

A substantial amount of work certainly needs to be done to organize content into digestible clips and learning-paths in order to make this a reality. This work should not be creators' responsibility — creators should be free to focus on creating their best content, without worry of length, sectioning, or learning-paths.

We designed Discite to automate this process. Discite will:

• Automatically process a video and extract relevant topics and subtopics for each period, using machine learning models that take into account both visual and audio (transcript) context.
  Read more in Docs > Architecture > Machine Learning.
• Split the video into atomic clips, taking care to not section in the middle of an explanation.
  Read more in Docs > Architecture > Clipping.
• Tag each clip with the relevant topic, subtopic and fine-grained title.

System Design and Architecture

Discite is mobile-first. We have various components that work together to make this a reality.

• Mobile App: The mobile app is the primary interface for users to consume content. It is built using Swift.
  Read more in Docs > Architecture > Frontend.
• Backend: We use Node.js and Express for out backend, which provides a reasonable API between our user-facing app and our data.
  Read more in Docs > Architecture > Backend.
• Database: We use MongoDB to store our data, with well-designed and optimized schemas to ensure efficiency and scalability.
  Read more in Docs > Architecture > Database.
• Video Streaming: We use HLS to stream videos to our users in real-time, with videos hosted on Amazon S3.
  Read more in Docs > Architecture > Streaming.
• Machine Learning: We use models built in PyTorch, Scikit-Learn, and Gensim to process videos and extract topics and subtopics.
  Read more in Docs > Architecture > Machine Learning.
• Clipping: We use FFmpeg to split videos into atomic clips. We compute medians and means of topics to determine when to split a video.
  Read more in Docs > Architecture > Clipping.

We loved building Discite, and we hope you love using it too!

If you have any feedback, please complete our feedback form.
You can also share with friends using the icons below.

Thank you for checking us out!