you2idea@video:~$ watch QIA9Ry42Lbs [13:00]
// transcript — 147 segments
0:03 This is what most founders are doing right now. >> They're all building the same open AI
0:10 rappers and competing for demand. But instead, this is what you should be
0:16 doing. Build niche AI solutions for small but carefully targeted group of
0:21 people. When you say AI today, most of you think about Cad GBT, Gemini, and
0:25 Clock. But in reality, there are thousands, in fact millions of small,
0:30 super specialized AI models available to use, and many of them are free and open
0:34 source. By fine-tuning, combining, and chaining specific models, you can solve
0:38 many different niche problems for a small group of users and turn it in to a
0:44 thriving business. Sounds complex? Well, you need a bit of technical
0:47 understanding, but in reality, you can put a SAS like this together using just
0:51 a few tools, and you can do it without having to code. In this video, I'll show
0:55 you how to approach this step by step. So, here's the problem I set out to
0:59 solve. My team and I produce a lot of video content, but I can't stand generic
1:04 stock videos like the ones you find on pixels. On the other hand, shooting
1:09 custom B-roll for every video is really timeconuming. So, I wanted a SAS that
1:13 could create highquality AI generated B-roll clips with me in it simply giving
1:17 a description in natural language. And that's exactly what we're going to build
1:21 in this video. a small SAS tool where you can upload images of yourself, use
1:25 it to create an AI character of yourself, then use that character to
1:31 create AI generated B-roll clips. So, this SAS is solving a problem I have
1:34 myself and I know a few other people who have the same problem, but it's not a
1:39 big market and that's the whole point. We're building a niche solution for a
1:43 small group of people here. I suggest you do the same. Use the principles I'm
1:47 about to show you, but find a problem you have and try to solve that for
1:53 yourself first. Let's get started. Let's do a quick architectural breakdown. A
1:57 product like this will have three layers. The first layer is a front end.
2:03 This is the part the user will be using. We'll use lovable to vibe code this
2:09 part. The second layer is a back end. This is where we'll store items in a
2:13 database such as characters and B-roll clips. We'll use Superbase for this
2:19 part. Finally, the last layer is our AI engine. We need a place to host and run
2:25 our AI models. We'll use replicate for this part. So, let's get started with
2:30 the front end. Head over to lovable.dev and we'll give it a very highlevel
2:35 prompt to begin with. No functionality yet, just the UI. And after running for a bit, it will set
2:43 up a new project for us. Awesome. We now have the UI for uploading images and
2:48 creating a new character. We also have some basic UI for generating a clip and a basic history
2:54 page. I think the design is pretty on point. I'll take it. Now, as mentioned, we need
3:00 both a front end and a back end. And for the back end, we're going to use
3:04 Superbase. And here's the great news. Lovable now has a native superbase
3:08 integration, what they call Lovable Cloud. So we can now handle both the
3:12 front end and the back end through Lovable, which makes things a whole lot
3:16 easier. So let's just go ahead and click connect Lovable Cloud.
3:22 Then enable cloud here and then allow Lovable to set it all up.
3:28 And now if we click the cloud button up here, we'll see that we have all these
3:32 things available in our project. These are all the things that make up our back
3:38 end. Awesome. So far so good. Now, the way this SAS works is by uploading a
3:42 bunch of images in a zip file in order to create a character. Now that we have
3:46 the backend set up, let's have Lovable create this functionality for us.
3:57 [Music] And here we'll see Lovable ask for permission to update some stuff on the
4:03 cloud including a database table and storage. We'll allow this and all of our backend
4:10 stuff will be set up. Lovable and Superbase comes with user authentication built in and it's very
4:15 easy to set up. So you can do that now if you want to or for building the PC
4:19 you can simply tell Lovable to just assume a mock user for now. That's what
4:24 I'll do. Perfect. Now, after a bit of working, we should have a form that we
4:29 can use to create a character Okay, seems to work. Let's just verify.
4:38 If you go to the cloud section here, then go to storage, and then there
4:44 should be a folder for our mock user. And there we go. The zip was uploaded.
4:48 Perfect. Now, this is probably a good time to make a quick note that with these types
4:53 of AI coding tools, things don't always go as smoothly as they just did here. In
4:58 many cases, something won't work, and you'll be going back and forth with
5:01 Lovable for a bit until things play nicely. But it's totally normal and not
5:05 very different from how things would flow if you were working with a team of
5:09 human developers. So, just keep in mind that this is an iterative process and
5:13 some patience is required. Now, let's move on to the fun part. We need to set
5:18 up the chain of AI models we'll use for this SAS. Let's get an overview. When a
5:22 new character is created, we want to trigger a training job. This will use
5:26 the uploaded images to create a custom image generation model that can generate
5:30 new images with the character in them. This process is called fine-tuning.
5:34 Then, when a new B-roll clip is created, we chain two models. The custom image
5:39 generation model will generate an image. The image is then fed to a video
5:42 generation model and used as the starting frame for the clip.
5:47 We will do this on replicate. So head over to replicate.com and create a new
5:51 account. Here we'll use the fastflux trainer model to fine-tune our custom
5:55 image generation model. And we'll use clingv 2.1 to turn an image into a
6:01 video. And I know all this might sound complex, but don't worry cuz lovable
6:05 already knows how to use the replicate library and intuitively does a great job
6:10 in wiring this together. And I'll show you how. in our replicate dashboard.
6:19 And back in the cloud section of Lovable, let's go to secrets and add
6:31 We're also going to take the replicate username, which is the one you'll see up
6:36 here, and give that to lovable under secrets, Good. Let's ask Lovable to use the
6:51 replicate API to start training our And as you can see, we're given Lovable
7:04 a few technical details here. And again, Lovable will probably ask you
7:08 to allow it to create some database tables and update security policies and
7:13 so on. Just let it do its thing and it will set up everything for you. And if
7:17 you don't know any technical stuff like this, don't worry. With Lovable, you'll
7:20 get there eventually. It would just take longer and require more patience. So,
7:25 giving these AI coding tools some technical descriptions really does help
7:29 a lot. So, try to do that if and whenever you can. Awesome. Now hopefully
7:33 maybe with a bit of back and forth you'll have something like this. So we
7:47 and it will start training that character on your images and replicate under the trainings tab.
7:54 We should now see our flux trainer running. After a few minutes, the character will
8:01 finish and the UI will update like this. Awesome. And in Replicate, we can verify
8:05 this by clicking over to the models tab. Here we should see our new custom model.
8:11 Perfect. The first part of our SAS is now working. Now, if we go to the
8:15 generate B-roll page, we currently have a form that looks nice but does nothing.
8:20 This is the next part we need to tell Lovable to implement. So, first we want
8:24 to get the configuration for the image model. And the easiest way to do this is
8:28 to simply test it out on replicate first. So we're going to add our test
8:43 And since we want a bit higher quality images, let's increase this to around
8:47 40. PNG All right, there we go. Looks good to me. So, let's click over here and we can
9:12 Perfect. Now, let's use this and give Lovable a big prompt here. We'll give it
9:17 the configurations here. And again, it's good to remind Lovable how replicates
9:21 API works so it can implement a proper polling mechanism. All right. Again, you might have to do a
9:28 little back and forth with Lovable if this step doesn't work from the get- go,
9:32 but you should get to a flow that works something like this.
9:37 We choose the character from the list. We can give it a style, though, I don't
9:47 actually think this does anything yet. And finally, let's run it.
9:55 And there we go. Perfect B-roll clip of our character, which is me in this case
10:00 walking in an office. Awesome, man. Just super cool. And if we go to replicate
10:04 and click the predictions tab, we should be able to see these two predictions
10:08 from the API. The one from the custom image and then the one from cling.
10:13 Looks like it's working perfectly. And one thing I absolutely love about
10:18 lovable is how good it is at filling in the gaps. We have toast messages,
10:22 spinners, various disabled and loading states, progress bars, and so on and so
10:27 on. I didn't ask for any of that. Lovable simply figured that we probably
10:31 would want that. Before AI, we'd have to manually sit and code all of that stuff,
10:35 and it would take forever. And if we would use a development agency, we would
10:39 have to explicitly ask for all of that, which for a lot of non techch founders
10:43 just isn't something they think about. So they'd get this super halfbaked
10:47 product back because that's what they ask for. Lovable does this extremely
10:51 well and I think it's such a cool experience. Finally, let's create the
10:55 history page. This one should be very straightforward since everything is set
11:04 And finally, you'll have a history page like this where you can see all your
11:11 And if we go to the cloud section in Lovable, we'll see that we now have a
11:16 database with two tables in it. We have a storage bucket for the zip
11:21 files and we have a bunch of these cloud functions. This is basically backend
11:26 code. Lovable created all of this for us and we didn't have to touch any code at
11:31 any point. Now, we also notice that we get this warning. This is because we
11:35 didn't implement user authentication earlier. So, if I publish this app,
11:39 everyone could create B-roll clips using my replicate account, which is obviously
11:43 a huge issue. So when you see this warning, you should make sure to resolve
11:48 all security issues before publishing your app. And obviously, it doesn't
11:53 necessarily need to stop here. In between these two models, we could have
11:57 added GPT5 to automatically enhance the prompts before feeding it into the
12:00 replicate models. And we could have added an upscaling step at the end to
12:05 get full 4K quality using a video upscaler model and just about a hundred
12:09 other things. Replicate already has models for a ton of different things.
12:13 But if you really want to go niche and it's not already on Replicate, you can
12:17 go to hawking phase and find an additional 2 million super specialized
12:22 AI models, fine-tune them, and upload them to your replicate account.
12:27 Everything that can be solved with AI is available here. And as I've mentioned a
12:31 few times now, in this video, I've shown the cut down happy path of building a
12:35 tool like this. In the real world, it's unlikely to go this smooth in the first
12:39 try. And typically you'll have to go back and forth with lovable for a bit.
12:43 So don't get discouraged if it doesn't spin up a SAS for you exactly how you
12:47 want it in the first try. And if you want to take a SAS like this to the next
12:51 level, you should layer the approach we've discussed in this video with two
12:55 additional layers of AI assisted coding which I cover in this video. So go watch
$

Get rich building niche AI SaaS (...not another ChatGPT wrapper)

[AI agents and automation][market research and validation][community building and audience growth][content creation and YouTube]
// description

Get full access to my SaaS portfolio - for a single one-time purchase. 📦 https://simonl.ink/founderstack Join my free SaaS newsletter - lightweight, easy-read, tips & insights. 📨 https://simonl.ink/newsletter --- Tools: - Lovable (https://lovable.dev) - Replicate (https://replicate.com) - Hugging Face (https://huggingface.co) --- Watch next: https://www.youtube.com/watch?v=IQqNRmoEa0s

now: 0:00
// tags
[AI agents and automation][market research and validation][community building and audience growth][content creation and YouTube]