Recast

Recast logo
Model architecture

Inspiration

We were inspired by our own problems with video streaming services, including buffering, poor quality, and throttling by ISPs. We decided to build a solution for this issue by tackling video compression.

What it does

Recast intelligently compresses videos by learning encodings from real video data. This allows for more efficient video streaming and storage in any domain.

How we built it

We built Recast using a Python backend, a React frontend, and AWS hosting. This included deep learning models trained with PyTorch, various AWS services, and a single page application built with React.

Challenges we ran into

One primary challenge we faced was designing a deep learning architecture specifically for this task. We ended up using a convolutional auto-encoder trained on a specific loss function for this problem after some experimentation. We also ran into the challenge of setting up the whole cloud infrastructure in a limited time; luckily, we managed to build an entirely cloud-native application.

Accomplishments that we're proud of

One of our teammates, Lui, is a premed student and a Biochemistry major. He had no experience with computer science or programming whatsoever prior to this weekend— but he was able to build his very own React app and maintain a GitHub repository. In fact, three of our team members had never attended a hackathon before, but we still managed to achieve our targets for Recast.

What we learned

We learned that it’s possible to build a domain-specific video compression model that also generalizes quite well. We also learned the intricacies of bringing multiple services together in a cloud environment. The process of shifting from local to cloud-hosted applications presented itself with many integration and coordination considerations, and we were able to learn how to best architect cloud infrastructure for our needs.

What's next for Recast

Future goals for Recast are training on larger datasets to provide models with better generalization. We also want to test Recast in a variety of domains, including healthcare and education.