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

Full transcript

1. 0:00This is what most founders are doing
2. 0:03right now.
3. 0:04>> They're all building the same open AI
4. 0:07rappers and competing for demand.
5. 0:10But instead, this is what you should be
6. 0:13doing. Build niche AI solutions for
7. 0:16small but carefully targeted group of
8. 0:18people. When you say AI today, most of
9. 0:21you think about Cad GBT, Gemini, and
10. 0:23Clock. But in reality, there are
11. 0:25thousands, in fact millions of small,
12. 0:27super specialized AI models available to
13. 0:30use, and many of them are free and open
14. 0:32source. By fine-tuning, combining, and
15. 0:34chaining specific models, you can solve
16. 0:36many different niche problems for a
17. 0:38small group of users and turn it in to a
18. 0:41thriving business. Sounds complex? Well,
19. 0:44you need a bit of technical
20. 0:45understanding, but in reality, you can
21. 0:47put a SAS like this together using just
22. 0:49a few tools, and you can do it without
23. 0:51having to code. In this video, I'll show
24. 0:53you how to approach this step by step.
25. 0:55So, here's the problem I set out to
26. 0:57solve. My team and I produce a lot of
27. 0:59video content, but I can't stand generic
28. 1:02stock videos like the ones you find on
29. 1:04pixels. On the other hand, shooting
30. 1:06custom B-roll for every video is really
31. 1:09timeconuming. So, I wanted a SAS that
32. 1:11could create highquality AI generated
33. 1:13B-roll clips with me in it simply giving
34. 1:15a description in natural language. And
35. 1:17that's exactly what we're going to build
36. 1:19in this video. a small SAS tool where
37. 1:21you can upload images of yourself, use
38. 1:24it to create an AI character of
39. 1:25yourself, then use that character to
40. 1:28create AI generated B-roll clips. So,
41. 1:31this SAS is solving a problem I have
42. 1:32myself and I know a few other people who
43. 1:34have the same problem, but it's not a
44. 1:36big market and that's the whole point.
45. 1:38We're building a niche solution for a
46. 1:41small group of people here. I suggest
47. 1:43you do the same. Use the principles I'm
48. 1:45about to show you, but find a problem
49. 1:47you have and try to solve that for
50. 1:49yourself first. Let's get started. Let's
51. 1:53do a quick architectural breakdown. A
52. 1:56product like this will have three
53. 1:57layers. The first layer is a front end.
54. 2:01This is the part the user will be using.
55. 2:03We'll use lovable to vibe code this
56. 2:05part. The second layer is a back end.
57. 2:09This is where we'll store items in a
58. 2:11database such as characters and B-roll
59. 2:13clips. We'll use Superbase for this
60. 2:16part. Finally, the last layer is our AI
61. 2:19engine. We need a place to host and run
62. 2:22our AI models. We'll use replicate for
63. 2:25this part. So, let's get started with
64. 2:27the front end. Head over to lovable.dev
65. 2:30and we'll give it a very highlevel
66. 2:32prompt to begin with. No functionality
67. 2:35yet, just the UI.
68. 2:38And after running for a bit, it will set
69. 2:40up a new project for us. Awesome. We now
70. 2:43have the UI for uploading images and
71. 2:45creating a new character.
72. 2:48We also have some basic UI for
73. 2:49generating a clip and a basic history
74. 2:52page.
75. 2:54I think the design is pretty on point.
76. 2:56I'll take it. Now, as mentioned, we need
77. 2:58both a front end and a back end. And for
78. 3:00the back end, we're going to use
79. 3:02Superbase. And here's the great news.
80. 3:04Lovable now has a native superbase
81. 3:06integration, what they call Lovable
82. 3:08Cloud. So we can now handle both the
83. 3:10front end and the back end through
84. 3:12Lovable, which makes things a whole lot
85. 3:14easier. So let's just go ahead and click
86. 3:16connect Lovable Cloud.
87. 3:19Then enable cloud here
88. 3:22and then allow Lovable to set it all up.
89. 3:26And now if we click the cloud button up
90. 3:28here, we'll see that we have all these
91. 3:30things available in our project. These
92. 3:32are all the things that make up our back
93. 3:34end. Awesome. So far so good. Now, the
94. 3:38way this SAS works is by uploading a
95. 3:40bunch of images in a zip file in order
96. 3:42to create a character. Now that we have
97. 3:44the backend set up, let's have Lovable
98. 3:46create this functionality for us.
99. 3:50We're going to give it this prompt.
100. 3:54[Music]
101. 3:57And here we'll see Lovable ask for
102. 3:59permission to update some stuff on the
103. 4:01cloud including a database table and
104. 4:03storage.
105. 4:05We'll allow this and all of our backend
106. 4:07stuff will be set up.
107. 4:10Lovable and Superbase comes with user
108. 4:12authentication built in and it's very
109. 4:13easy to set up. So you can do that now
110. 4:15if you want to or for building the PC
111. 4:18you can simply tell Lovable to just
112. 4:19assume a mock user for now. That's what
113. 4:22I'll do. Perfect. Now, after a bit of
114. 4:24working, we should have a form that we
115. 4:26can use to create a character
116. 4:29and upload a SIP.
117. 4:33Okay, seems to work. Let's just verify.
118. 4:36If you go to the cloud section here,
119. 4:38then go to storage, and then there
120. 4:40should be a folder for our mock user.
121. 4:44And there we go. The zip was uploaded.
122. 4:47Perfect.
123. 4:48Now, this is probably a good time to
124. 4:49make a quick note that with these types
125. 4:51of AI coding tools, things don't always
126. 4:53go as smoothly as they just did here. In
127. 4:56many cases, something won't work, and
128. 4:58you'll be going back and forth with
129. 4:59Lovable for a bit until things play
130. 5:01nicely. But it's totally normal and not
131. 5:03very different from how things would
132. 5:05flow if you were working with a team of
133. 5:07human developers. So, just keep in mind
134. 5:09that this is an iterative process and
135. 5:10some patience is required. Now, let's
136. 5:13move on to the fun part. We need to set
137. 5:15up the chain of AI models we'll use for
138. 5:18this SAS. Let's get an overview. When a
139. 5:20new character is created, we want to
140. 5:22trigger a training job. This will use
141. 5:24the uploaded images to create a custom
142. 5:26image generation model that can generate
143. 5:28new images with the character in them.
144. 5:30This process is called fine-tuning.
145. 5:32Then, when a new B-roll clip is created,
146. 5:34we chain two models. The custom image
147. 5:36generation model will generate an image.
148. 5:38The image is then fed to a video
149. 5:40generation model and used as the
150. 5:42starting frame for the clip.
151. 5:44We will do this on replicate. So head
152. 5:47over to replicate.com and create a new
153. 5:49account. Here we'll use the fastflux
154. 5:51trainer model to fine-tune our custom
155. 5:53image generation model. And we'll use
156. 5:55clingv 2.1 to turn an image into a
157. 5:58video. And I know all this might sound
158. 6:01complex, but don't worry cuz lovable
159. 6:03already knows how to use the replicate
160. 6:05library and intuitively does a great job
161. 6:08in wiring this together. And I'll show
162. 6:10you how. in our replicate dashboard.
163. 6:12Let's go get the API key.
164. 6:17And back in the cloud section of
165. 6:19Lovable, let's go to secrets and add
166. 6:22that API key here.
167. 6:29We're also going to take the replicate
168. 6:31username, which is the one you'll see up
169. 6:33here,
170. 6:36and give that to lovable under secrets,
171. 6:38too.
172. 6:47Good. Let's ask Lovable to use the
173. 6:49replicate API to start training our
174. 6:51model. This is the prompt we'll give it.
175. 6:58And as you can see, we're given Lovable
176. 7:00a few technical details here.
177. 7:04And again, Lovable will probably ask you
178. 7:07to allow it to create some database
179. 7:08tables and update security policies and
180. 7:10so on. Just let it do its thing and it
181. 7:13will set up everything for you. And if
182. 7:15you don't know any technical stuff like
183. 7:17this, don't worry. With Lovable, you'll
184. 7:19get there eventually. It would just take
185. 7:20longer and require more patience. So,
186. 7:23giving these AI coding tools some
187. 7:25technical descriptions really does help
188. 7:27a lot. So, try to do that if and
189. 7:29whenever you can. Awesome. Now hopefully
190. 7:32maybe with a bit of back and forth
191. 7:33you'll have something like this. So we
192. 7:36can now create a character
193. 7:45and it will start training that
194. 7:47character on your images
195. 7:49and replicate under the trainings tab.
196. 7:52We should now see our flux trainer
197. 7:54running.
198. 7:56After a few minutes, the character will
199. 7:57finish and the UI will update like this.
200. 8:00Awesome. And in Replicate, we can verify
201. 8:02this by clicking over to the models tab.
202. 8:05Here we should see our new custom model.
203. 8:08Perfect. The first part of our SAS is
204. 8:10now working. Now, if we go to the
205. 8:13generate B-roll page, we currently have
206. 8:15a form that looks nice but does nothing.
207. 8:18This is the next part we need to tell
208. 8:20Lovable to implement. So, first we want
209. 8:22to get the configuration for the image
210. 8:24model. And the easiest way to do this is
211. 8:26to simply test it out on replicate
212. 8:28first. So we're going to add our test
213. 8:31prompt here.
214. 8:36We'll use a 169 aspect ratio.
215. 8:42And since we want a bit higher quality
216. 8:43images, let's increase this to around
217. 8:4540.
218. 8:47PNG
219. 8:50full output quality. And let's run it.
220. 8:58All right, there we go. Looks good to
221. 9:00me. So, let's click over here and we can
222. 9:03simply copy all these configurations.
223. 9:09Perfect. Now, let's use this and give
224. 9:12Lovable a big prompt here. We'll give it
225. 9:14the configurations here. And again, it's
226. 9:17good to remind Lovable how replicates
227. 9:19API works so it can implement a proper
228. 9:21polling mechanism.
229. 9:24All right. Again, you might have to do a
230. 9:27little back and forth with Lovable if
231. 9:28this step doesn't work from the get- go,
232. 9:30but you should get to a flow that works
233. 9:32something like this.
234. 9:35We choose the character from the list.
235. 9:37We give it a prompt.
236. 9:42We can give it a style, though, I don't
237. 9:44actually think this does anything yet.
238. 9:47And finally, let's run it.
239. 9:51And there we go. Perfect B-roll clip of
240. 9:55our character, which is me in this case
241. 9:57walking in an office. Awesome, man. Just
242. 10:00super cool. And if we go to replicate
243. 10:02and click the predictions tab, we should
244. 10:04be able to see these two predictions
245. 10:06from the API. The one from the custom
246. 10:08image and then the one from cling.
247. 10:11Looks like it's working perfectly. And
248. 10:13one thing I absolutely love about
249. 10:15lovable is how good it is at filling in
250. 10:18the gaps. We have toast messages,
251. 10:20spinners, various disabled and loading
252. 10:22states, progress bars, and so on and so
253. 10:25on. I didn't ask for any of that.
254. 10:27Lovable simply figured that we probably
255. 10:29would want that. Before AI, we'd have to
256. 10:31manually sit and code all of that stuff,
257. 10:34and it would take forever. And if we
258. 10:35would use a development agency, we would
259. 10:37have to explicitly ask for all of that,
260. 10:39which for a lot of non techch founders
261. 10:41just isn't something they think about.
262. 10:43So they'd get this super halfbaked
263. 10:45product back because that's what they
264. 10:47ask for. Lovable does this extremely
265. 10:49well and I think it's such a cool
266. 10:51experience. Finally, let's create the
267. 10:53history page. This one should be very
268. 10:55straightforward since everything is set
269. 10:56up now.
270. 11:02And finally, you'll have a history page
271. 11:04like this where you can see all your
272. 11:06previous clips.
273. 11:10And if we go to the cloud section in
274. 11:11Lovable, we'll see that we now have a
275. 11:13database with two tables in it.
276. 11:16We have a storage bucket for the zip
277. 11:18files and we have a bunch of these cloud
278. 11:21functions. This is basically backend
279. 11:23code. Lovable created all of this for us
280. 11:26and we didn't have to touch any code at
281. 11:28any point. Now, we also notice that we
282. 11:31get this warning. This is because we
283. 11:33didn't implement user authentication
284. 11:35earlier. So, if I publish this app,
285. 11:37everyone could create B-roll clips using
286. 11:39my replicate account, which is obviously
287. 11:41a huge issue. So when you see this
288. 11:43warning, you should make sure to resolve
289. 11:46all security issues before publishing
290. 11:48your app. And obviously, it doesn't
291. 11:51necessarily need to stop here. In
292. 11:53between these two models, we could have
293. 11:55added GPT5 to automatically enhance the
294. 11:57prompts before feeding it into the
295. 11:59replicate models. And we could have
296. 12:00added an upscaling step at the end to
297. 12:02get full 4K quality using a video
298. 12:05upscaler model and just about a hundred
299. 12:07other things. Replicate already has
300. 12:09models for a ton of different things.
301. 12:11But if you really want to go niche and
302. 12:13it's not already on Replicate, you can
303. 12:15go to hawking phase and find an
304. 12:17additional 2 million super specialized
305. 12:20AI models, fine-tune them, and upload
306. 12:22them to your replicate account.
307. 12:25Everything that can be solved with AI is
308. 12:27available here. And as I've mentioned a
309. 12:29few times now, in this video, I've shown
310. 12:31the cut down happy path of building a
311. 12:33tool like this. In the real world, it's
312. 12:35unlikely to go this smooth in the first
313. 12:37try. And typically you'll have to go
314. 12:39back and forth with lovable for a bit.
315. 12:41So don't get discouraged if it doesn't
316. 12:43spin up a SAS for you exactly how you
317. 12:45want it in the first try. And if you
318. 12:47want to take a SAS like this to the next
319. 12:49level, you should layer the approach
320. 12:51we've discussed in this video with two
321. 12:53additional layers of AI assisted coding
322. 12:55which I cover in this video. So go watch
323. 12:58this one next.