I've created Livi! It is an internal tool for me and my team so we can encode complex images that sometimes have plenty of detail and transparency into optimized AVIF files for our websites.
Backend is Rust and Frontend is SwiftUI. At some point I'll make a Libadwaita frontend for a Linux release. Given my knowledge in Swift and Rust is pretty limited, it was an interesting project to learn the strenghts and limits of LLMs. I've learned quite a lot with it. Most useful lesson is that you might not necessarily need to know the specifics of a language, but you do have to have your common sense skills sound and clear, and how to code architecture a larger project, with refactors here and there, performance optimizations, multithreading, queue, cache and logs, and so on.
Maybe needless to say, but it wasn't easy. The produced code needed to be inspected constantly, and bugfixing, testing, handling edge cases required tons of prompting and guidance. The comparison features, such as pinch to zoom, keep the zoom and image positions while switching between the different generated images, handling exporting, all these features were loaded with intricacies. So far, glad that I was able to produce this.
As a fun fact, now that AVIFENC supports tune=iq, we don't have to mess with specific encoder settings and find the proper quality number anymore, but still, learned a lot from it
https://github.com/MARTYR-X-LTD/livi
Backend is Rust and Frontend is SwiftUI. At some point I'll make a Libadwaita frontend for a Linux release. Given my knowledge in Swift and Rust is pretty limited, it was an interesting project to learn the strenghts and limits of LLMs. I've learned quite a lot with it. Most useful lesson is that you might not necessarily need to know the specifics of a language, but you do have to have your common sense skills sound and clear, and how to code architecture a larger project, with refactors here and there, performance optimizations, multithreading, queue, cache and logs, and so on.
Maybe needless to say, but it wasn't easy. The produced code needed to be inspected constantly, and bugfixing, testing, handling edge cases required tons of prompting and guidance. The comparison features, such as pinch to zoom, keep the zoom and image positions while switching between the different generated images, handling exporting, all these features were loaded with intricacies. So far, glad that I was able to produce this.
As a fun fact, now that AVIFENC supports tune=iq, we don't have to mess with specific encoder settings and find the proper quality number anymore, but still, learned a lot from it