Here, you can check out my clip aesthetic score selection at work, it picked the best frame. Kelly actually looks good in this frame. I barely smile, so I can't take a photo. Hey, there's a feature right here, taking photos of people who don't smile well. I'll make them smile even if they don't. I wanted to show a demo of this thing, but I ran into some weird flakes, so I had to go restart the demo. Let me take one last look at the video uploaded. It's just the video uploaded. So there's a generic post process, which could include other activities beyond encoding, right? Yeah, right when it's encoded, then you want to send it off to another message queue that does all the trimming and stuff like that. Or after it gets compressed. It's like... This is our server. Yeah, that's like the Amazon EC2. And so this is what the front end sends to. I think there is some stamp in between here. Probably before this or somewhere, we should do deduplication, frame augmentation. There's a small video process. Do we want to do that with after encoding or before encoding? I'm not a pro at that part. I think it comes after, right? I can now give a painstakingly detailed description of what actually happens because I just implemented it. The hardware encoder on this device outputs a stream of bytes, which form a MPEG standard H.265 block. And this block doesn't have a strong understanding of what frames are anymore. Because it's variable length encoded based on how much compression is applied to each frame. So when you have a finished video, that's the video. It gets sent to the back end. So at object store raw video, you receive a compressed set of frames. Then at encoding, you would first decode all of the frames. Then do all of your media processing, like keyframe selection. Oh, decoding. You first decode it. Then you can do keyframe selection, best frame identification, moment extraction, whatever. Then at the very end, you can optionally re-encode the video with this extra information if you care. So why do you encode and decode? The encoding makes the video 500 times smaller for uploading. Otherwise, you can't upload it fast enough. The encoding is done on device? Yeah, it's done in hardware on device. The soul of this device is hardware H.265 encoder. So is this encoder more standard encoder or for different video or different query? 100% standard. It's just H.265. When we want to do some video processing, like run some diffusion on top of some frames or some part of the videos, I don't think those can operate on H.265. No, you decode in the back end. This is typically like a user goes on YouTube and uploads a video.