Hassan El Mghari built every project he discusses over a weekend, using simple ideas, existing AI APIs, and fast iteration.
2
Small AI apps can attract large audiences because most people cannot build them themselves, even when they look like simple ChatGPT wrappers.
3
Free, open-source projects with polished interfaces can grow through search, word of mouth, sharing, and infrastructure credits from companies.
Summary
Hassan El Mghari walks through the side projects he built during weekends, including QR code generation, article summaries, AI Commit, photo restoration, and RoomGPT. His projects reached millions of visitors despite using relatively simple product ideas and off-the-shelf models. He argues that builders should launch narrow MVPs quickly, avoid fine-tuning and custom models at the start, use newly released models when they fit the problem, and iterate after seeing real usage. Open source helped his projects spread, while keeping them free required sponsorships and usage limits. He also spends most of his development time on the interface, since users judge the product before they judge the model. His final advice is about working in focused sprints, then resting, and continuing to build long enough to improve through practice.
Simple AI apps can reach people who cannot build them themselves
El Mghari says many of his projects looked too simple to attract users, including apps that were little more than ChatGPT wrappers. He argues that AI engineers often underestimate the majority of users, since many people have not used ChatGPT and cannot quickly build a small AI tool themselves. His Twitter bio generator reached about 200,000 visitors, including 100,000 in one weekend, before he had to shut it down temporarily after hitting his OpenAI bill. The lesson is that a narrow product can be useful even when its implementation is straightforward.
Traffic came mostly from search and sharing rather than Twitter followers
Although El Mghari has an audience on Twitter, he says less than 5% of the traffic across his projects came from his own Twitter account. He attributes the rest to word of mouth, Google, SEO, and other influencers sharing the products. RoomGPT reached about six million site visitors and a little over two million registered users, with most traffic coming directly from Google. Open-source code also encouraged developers to share the project and contribute pull requests. His experience challenges the idea that a large personal following is required before launching a side project.
A single model call can turn a small idea into a useful product
The photo restoration project takes an uploaded image, sends it to GFPGAN, and returns a restored version. El Mghari describes the core implementation as one API call to the model. The app still reached about 250,000 monthly users, especially in India and Indonesia, where he says lower-quality phone cameras created a strong use case. After the project went viral, he added TensorFlow.js to scan uploaded images for inappropriate content before processing them. The example shows how a simple pipeline can support a real consumer tool when the result is visible and immediately useful.
ControlNet made RoomGPT better at preserving a room's original structure
RoomGPT lets users upload a room and choose design themes, then generates redesigned versions that preserve the room's layout. El Mghari had seen Stable Diffusion produce redesigned rooms with changed dimensions and depth. When ControlNet appeared, he noticed that it maintained the original structure much better and built a product around that capability. He launched quickly and posted updates repeatedly. The project reached 10,000 users in 12 hours, 30,000 in a day, 90,000 after two days, and 270,000 after three days. He credits early production use of a newly released model as one reason it spread.
El Mghari advises builders not to fine-tune models or build their own models when the goal is to launch an MVP over a weekend. He says a project should be narrow enough to describe in five words. RoomGPT prompted machine learning engineers to ask about training data, parameters, and model design, but he had simply used an off-the-shelf API. His approach is to reduce the scope, launch in one or two weekends, and then decide whether to continue based on actual usage. If the product fails, the builder can move to another idea. If it works, the product can grow afterward.
Open source and free access can create distribution, but they need a funding plan
El Mghari says open source gives people a reason to learn from a project, share it, and submit pull requests. It also helped him gain about 25,000 Twitter followers from developers interested in his work. Free access is harder because AI workloads cost money. He sometimes asked model and infrastructure companies for credits in exchange for a mention in the app footer or README. For builders without an audience, he suggests making a high-quality open-source project, setting a limit such as $50, and contacting the company after the project gains attention. He names Replicate, Vercel, and Neon among companies that helped support his projects.
The interface takes most of the work because users judge what they can see
El Mghari says he spends about 80% of his time on the UI, even for AI products. He believes people will not use a product if it looks poor, and he does not rely on being a trained designer. Instead, he studies several websites, takes elements he likes from each, and combines them into a usable design. He also demonstrates v0, a Vercel tool that generates React UI from prompts, lets users inspect the code, fork designs, edit selected page elements, and stream updated components into the code view. The tool helps him prototype interfaces faster.
Weekend sprints work because focused effort is followed by time away
El Mghari says he does not work every day despite building many projects. He usually spends one weekend in an intense sprint, with about 12 hours on Saturday and 12 hours on Sunday, then spends the next two or three weekends relaxing, watching Netflix, and seeing friends. He describes this as roughly one project weekend each month. He acknowledges that people with children, marriages, or other responsibilities will need a different schedule. The point is to find a repeatable pattern that fits personal circumstances instead of copying his exact hours.
Building skill requires continuing past the first few projects
When people ask why their projects take so long, El Mghari often learns that the project is only their second. He compares this with going to the gym twice and expecting visible muscles. Coding and side-project work improve through repeated practice over months. He says he is not a professional engineer and does not write code for most of his job, having learned to code only a few years earlier. His conclusion is that almost anyone can build these products, but they have to put in enough hours and keep shipping projects to see progress.
"You can't go to the gym for the second time ever and then look down and be like, where are my biceps?"21:05
Who should watch
You have a weekend project idea and need a narrower MVP that can use existing models instead of custom training.
Your AI app works technically, but you need practical advice on distribution, open source, search traffic, and interface design.
You are waiting for a large audience or a sophisticated model before launching, and want examples of products built quickly without either.
