Agentic video editing pipeline

So my name is, the CTO and cofounder of Flashcut, and I'm going to talk about so Slacko is a, AI video assistant. And today, I'm gonna talk about the video intelligence feature. So in the room, does anyone know the streamer, speed, or does anyone want to travel CTO? Teams this is, just last week, I AWS, like, I was looking at the the video for for the speed and Chunk Teams, and I was like, okay. What if I want to look at, a section in the video where there's a there's a train that passed through, like, an apartment building? I think it's, like yeah. It it is kind of possible on YouTube, but, like, what if I want to do this more AI Editing it? So, this is basically what's what I'm doing. So, let's say AI have a video and I want to understand it. Here, the the 1st Dev is I want to standardize the video. If it's, like, 10 80 p, I don't really need to do that time. I need to, like, down sample it and have it at, up, like, 82 to 4. So, 1st, I have transcoded, then upload it to the cloud. So, Google Google Cloud Storage. And then, in order to make it usable, I want to cut it into small pieces. And I want tool segment it properly. I don't want to cut it where there's someone who's, AI, like, someone who can talk in the middle. So I there's like, I uses the f m n type, for that. Hat once I have these pieces, I feed it into the, Jevirah. So this is, the conference. This is the output schema. Lead. So I Rue Jet. I use Jet. Right? Yes. Okay. Yeah. So, I use the GenAI CTO. It used Gemini Flashcut, and I think this AWS 3.3 0.1, but now I have to update to 3.5. So thank you for the reminder. Better yeah. So, once AI get once I get all the pieces, I guess, the file output Tullo, could be, like, the, like, Agent type, what what AI is there. Yeah, like, the different person and some, like, deal on specific level that's in my my usage. Then put all of this, and embedded into, like, like a 3 72 dimension and put City on, Postgres, Red g vector. And finally, I can search all of this, the same the same thing. So, embed the query into Gemini embedding 1 and use the, like, Postgres PG better, AI. I think I'm not exactly sure. But there there's a there's there's a way CTO query that. And then I get the results, and I can I can see it? Let me do a demo. So this is a talk between, it's like AI it's like 1 of these things I show. So, this is, like, 4 minutes, and I want to like, there's a image of a guitar. So, I want to do for guitar. Yeah. So so this will, so from the edges. But, yeah, so in the transcript, there isn't any mention of guitar, but there is a photo point. So this is where you see guitar again. Yeah. So this is where we are API, and, you know, I'll be I'll be at a GPT. So, if you have any ingestion AI my implementation details, what's the next step, which is you wanna try tool Demo? Come come come talk to me. Question? So I I've also been playing around with this problem. I've tried Gemini online at AI. I tool, like, 2026 has some capabilities Easy processing things. But AI have feedback given on that, or have you mostly done the slicing and use the Gemini DLMs for for your project? I think it's AI, I'll be honest. I haven't tried it. It's it's worth AI. I mean, my CTO feeling said that, because Video, City has a lot of information. So I don't know if, that could work, but I'm happy to API Learn. No. No. What There's limitations if I'm really curious. Like, Web you talk to them online, AI the only time it actually was the time. So I was just curious if you'd explore that avenue and found something else going on by the way. Thank you. Oh, just to be clear, are you saying, like, taking, like, a 32nd video and putting it through Gemini live? Or, like, how to Gemini Live and looking at the Video? Gemini Live showing at the Video is actually more. They'll transfer this other modules since you're just getting to it. But Yeah. I was just I wanted to share notes. That's good. Okay. Thank you. So I was wondering if we could apply this in the Health care system. We're we're working right now, you know, on cancer, AI, brain cell tumors. And if we are to run, let's say, a scan, a brain scan, and we could have the model learn all the AI of different cancer cells, Could it identify in that sense? Just like a video when you look at it. Could it potentially help to look for something? Like, would it prompt or search if you did? Looking for a specific or, you know, having it detect something? I'm just trying to think out loud because this is where we're at, and I'm looking for a solution. That's a good question. Well, maybe if we we can take some frame out of that video and compare it to what the, like, what the brain tumor Ray looks for looks like, then compare it to. I think it could could pause. We're part of the reason. Yeah. So so all of this, very honestly, that's that's Hostile, but there's, just Staff a file for it. I'll I'll get it. There's Tullo, a district. There's a lot of competitor out there that does, like, have 12 labs that is, like, video understanding. I would say, I better than what I'm doing. So I will if you wanna do this, I suggest you go from there. But this is so this is mostly, like, tweak to file a case of what this like, how can how can, video by professional video editors use this so this is Forge video, like, professional video editors. So if you want to integrate with, like, audio and other, like, professional software, they can this is a step before. But, so the video understanding to help them, to, like, look at a different part of the video and see where I want to click the video. And how do I I put it it it it also, like, put it together to, like, automate the. Hi. I, for demo. What I will might also as a outside of tech channels, like, video editing. Oh. For, like, the and stuff like that. Well, the 4 key points for that segment of the industry is, like, compute requirements. Most editors use, like, have, like, long lessons of stuff, like, on our own platform. So can this be, like, run local? Usually, we have a high part between 2nd question. Can this be scaled across multiple beats, not just 1 long 5480. Do we have 3, 4, 5, 6 different, AI, video, Survives? Being able to, like, orchestrate between them has been AI of, like, pain point for Forge Right. Yeah. That's a Ray question. So the 1st point, this this runs on the cloud, so there's nothing Web can run locally. I think there's 1 so the only thing is we we're transcoding the video, which uses your your, like, your. But if it's like if if it is more than f 1, then it should be it should be. The so so the 2nd question for, like, if we have, like, multiple videos. Yes. So the target is, let's say you have, like, a main a row and you have, like, multiple b row that wants to use us, like, introduction and stuff. We we will have, like well, currently, it doesn't, but, the goal is to be able to have, like, have, like, multiple tasks, and you can you can have, like, the the main speaker and have, like, different, different video. So, like, 2 video AI, 1 is a roll, the other is, AI some b roll. And for the audio, the main audio and potentially have some, execution for the song. And then we can we have the export feature to, prepare or have API afterwards. Thank you. Oh, so there's a there's an AI chat. Here, you can Ray, like, hey. I hope, like, the log Code, and it will guide you to yeah. So thank you so much.