AI Engineer Summit 2023
The Weekend AI Engineer
Loading slides…
Read the talk
The Weekend AI Engineer
Hassan El Mghari’s weekend projects show how narrow workflows, off-the-shelf models and careful interfaces can turn small experiments into widely used AI applications.
From a talk by Hassan El Mghari
Before you start: Basic familiarity with web applications, APIs and Git is helpful; no model-training experience is required.
What can you build in a weekend that people will use?
A weekend gives you enough time to solve one problem, but very little room to build an entire platform. Hassan El Mghari’s projects explore that constraint: small applications with a clear input, a useful model call and an output someone immediately understands. After about two years of consistently building side projects, his previous year’s 11 projects had attracted about 20,000 visitors. His next goal was to double that to 40,000.
Hassan reports that the portfolio instead reached over 8 million unique visitors, 20,000 GitHub stars and about 2.8 million signups. Every project he launched was built on a weekend. The complete comparison slide puts the earlier year beside the much larger AI-project portfolio.
Hassan makes his projects open source. Open source serves several purposes at once: people can learn from the implementation, share the project and contribute improvements. Some of the most helpful pull requests improve prompts beyond what Hassan initially wrote.
An existing Twitter audience helps with a launch, but it does not explain all subsequent discovery. Hassan attributes less than 5% of traffic across a stated 8.5 million portfolio visitors to his own Twitter account. He points instead to word of mouth, Google and SEO, and other people sharing the applications. That distinction matters when considering whether a small project needs a large personal audience before it can be useful.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From hundreds of image descriptions to one styled QR code
The first practical problem was a conference photo gallery. Several hundred images were ready to publish when Hassan’s CEO asked for alt text. Writing those descriptions manually would have been tedious, so Hassan found an image-to-text API and used it to generate them. He checked many of the descriptions, recalls fixing roughly two, and published the gallery. This was his entry into AI the preceding December: a model removed a substantial publishing chore while leaving a manageable review step.
The time savings made AI feel immediately useful. Hassan jokes about the contrast with the previous year’s Web3 hype, then concedes that Web3 has its place. After releasing the alt-text project as open source, he continued experimenting and built QRGPT with his friend Kevin at a hackathon.
QRGPT takes a destination and a visual prompt. In the demonstration, the destination is ai.engineer, and the preset describes a forest overlooking a mountain. The intended result is a QR code that links to the conference while incorporating that scenery instead of presenting only a black-and-white pattern. Hassan expects generation to take five or six seconds; this is an onstage expectation, not a measured latency result. The finished frame shows the mountain-and-forest code beside Download and Share controls.
Hassan reports about 8,000 visitors and about 8,000 generated QR codes. That was a satisfying launch, but QR-code creation is an occasional task. The next question was whether he could build something people would return to regularly.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Summaries, glasses and commit messages
A TechCrunch summarizer targeted a more repeatable habit: reading news. Adding summary to an article URL redirected the reader to Hassan’s site, where GPT-3.5 reduced the article to a few bullet points. The demonstration is a recording because, in Hassan’s account, TechCrunch’s lawyers emailed him asking him to take the application down. He complied, and its initial traction ended with the shutdown. A convenient workflow still depended on the publisher whose content it summarized.
Other ideas came directly from public requests. When Samina asked on Twitter for AI help choosing glasses, Hassan built a small recommender in a few hours. It accepted face shape, gender and optional context, then combined an LLM with the Amazon API to find glasses and provide purchase links. The scope stayed narrow: collect the information needed for one recommendation and make the result actionable.
Next, Theo suggested an application that automatically generated commit messages. Hassan’s CTO tagged him, and the result became AI Commits. Its input is work the developer has already done: stage changes, let the CLI analyze the Git diff, and review a proposed message before using it to commit. The documented command is aicommits:
bash
git add src/
git diff --cached
aicommits
Inspecting the staged diff makes the scope explicit: those are the changes the generated message should describe.
The CTO’s tag arrived at 7:53 PM on February 11, and Hassan replied with a working script less than two hours later. Only after it attracted attention did he clean it up and work out npm packaging, spending Monday morning turning the script into something other developers could install. Hassan estimates that over 30,000 developers were using AI Commits, with about 6,000 GitHub stars and 25 contributors at the time of the talk. The sequence was a small working tool first, then packaging and community investment after demand appeared.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A simple wrapper can be a complete product
The Twitter Bio Generator reduces another open-ended writing task to a small form. Enter personal context, choose a vibe and receive a streamed GPT-3.5 response. Hassan’s example combines being an engineer at Microsoft with loving volleyball; the result includes “spiking code bugs and volleyball balls.” The interface asks for only the information necessary to produce the requested artifact.
To an AI engineer, this can look too simple to matter. Anyone familiar with an LLM could write the prompt or build the wrapper. But the intended user is not necessarily an AI engineer—or even someone who has used ChatGPT. A narrow interface can make an existing capability accessible to people who would never assemble the workflow themselves.
Hassan reports about 200,000 visitors to the bio generator, including about 100,000 in a single weekend. The traffic also exposed a constraint that would recur across the projects: he temporarily shut it down after hitting his OpenAI bill. A successful free application can create an inference-cost problem almost immediately.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Restore a face, then handle the uploads
The move into image-to-image applications began with old family photographs. RestorePhotos aims to make blurry faces clearer. Hassan uploads a photograph of his father doing karate: the pose is striking, but the face is difficult to see. The application sends the image to GFPGAN, a face-restoration model, and displays the returned image. During the live demonstration, the result takes long enough that Hassan moves on and returns to it later.
Hassan reports about 250,000 people using RestorePhotos each month. He identifies India and Indonesia as major sources of users and suggests that lower-quality phone cameras help explain the demand; that is his interpretation of the audience. Popularity also brought inappropriate uploads. He added an image-safety check using TensorFlow.js before processing and published that functionality as a library. The application now had an operational responsibility beyond restoration: deciding which inputs to send onward.
Returning to the completed restoration, Hassan puts the original and result side by side and zooms in on his father’s face. He points out the clearer result, then emphasizes how little model orchestration was involved: one API call to GFPGAN, followed by displaying the output. The value comes from connecting the model to a personal task people care about.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The shared path from upload to result
Most of Hassan’s image applications share the same small architecture. Next.js handles both the interface and the backend boundary:
- An upload component sends the user’s image to cloud storage.
- The client sends the stored image URL to a Next.js API route.
- That route calls the machine-learning model—GFPGAN in the restoration example.
- The route returns the result to the client, which displays it.
Hassan compares the API route to a Lambda function. Cloud storage holds the image, the route coordinates the model request, and the browser presents the result. This separation lets the same application structure support multiple image-to-image tasks.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
RoomGPT: change the décor without losing the room
RoomGPT applies that image workflow to interior redesign. Hassan uploads a living-room photograph from the internet and selects themes. The generated variations change colors, couch styles, tables, backgrounds and tiles while retaining the recognizable structure of the room. Preserving that structure is essential: a beautiful image of a different space is less useful when the user wants ideas for their own room.
The motivation came from an earlier room-redesign application using Stable Diffusion. Hassan found that its outputs changed room dimensions and depth enough to lose the original space. ControlNet supplied the structural control he wanted. Technically, ControlNet adds spatial conditioning to pretrained diffusion models, including Stable Diffusion; it is not simply a replacement for Stable Diffusion. The useful comparison here is the room-redesign behavior Hassan observed:
| Approach | Room-redesign behavior |
|---|---|
| Earlier Stable Diffusion application | Dimensions and depth drifted |
| ControlNet-based RoomGPT | Décor varied while room structure remained recognizable |
That concrete improvement gave him a reason to build and launch an open-source application.
After launching on Twitter, Hassan kept posting updates rather than treating the first announcement as the entire launch. He considers a tweet’s useful lifetime roughly a day. As usage grew, he added testimonials and posted new milestones, joking about whether he had paid the people supplying them. He reports this progression:
| Time after launch | Reported users |
|---|---|
| 12 hours | 10,000 |
| One day | 30,000 |
| Two days | 90,000 |
| Three days | 270,000 |
Hassan attributes much of the response to being among the first to turn the newly available ControlNet capability into an application people could use.
Hassan reports about 6 million cumulative RoomGPT visitors and a little over 2 million people who registered and used it. The analytics view shows the large initial spike and the long decline that followed. He describes Google as the dominant traffic source and credits repeated sharing, including developers sharing the open-source project. Keeping access free also reduced the barrier to trying it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Use AI to build the interface, too
The model can help build the application as well as power its main feature. Hassan recommends GPT-4 for coding and points to the Vercel AI SDK and v0. The v0 demonstration is its early October 2023 UI-prototyping workflow: browse generated interfaces, inspect their code and iterate on a copy.
He opens an Apple Notes-like interface made by another user. The code is available to inspect or copy, and the interface can be forked much like a GitHub repository. Once forked, it becomes a starting point for further prompts rather than a fixed template.
The edit is deliberately local. Hassan selects a div and asks v0 to add three more notes and alternate their colors. He describes the selected region rerendering as React components stream through the Vercel AI SDK, with generated code accumulating in the code box. The sequence moves from a requested change to a visible result: generation continues, notes three through five appear, and the completed interface contains five notes.
The result can be copied as code or retrieved with a CLI command for further work. That makes the workflow useful for early interface prototyping: start from something close, target the region that needs changing, and bring the generated code into the application.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep the first release small enough to finish
Alongside coding assistants and UI generation, Hassan names Replicate, Hugging Face, Modal and Brev as tools that help builders move faster or work with models. For a weekend MVP, however, his recommendation is to avoid training a model or fine-tuning one. The scope of that advice is the initial, quickly launched application: use an existing capability to test whether the product is useful.
His five-word description heuristic forces the idea toward a single job. A platform that connects developers with clients, hosts portfolios and includes chat already contains several products’ worth of work. If the idea cannot fit into a weekend, remove features until an end-to-end MVP can.
RoomGPT prompted questions from machine-learning engineers about training, parameters, data collection and cleaning. Hassan’s answer was that he had used an off-the-shelf API. He also emphasizes watching new model releases: in his account, ControlNet had become available only a couple of days before the opportunity he pursued. The advantage was recognizing a newly solvable user problem and packaging the capability quickly.
Launch before committing to a larger build. A release after one or two weekends limits the effort at risk. If the project fails, pivot or move to another idea. If it gains traction, add features and invest further. The first release is a way to learn which project deserves the next weekend.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Pay for free access without ignoring the bill
Open source gives people reasons to learn from a project, share it and submit improvements. It can also build an audience around the work: Hassan reports gaining about 25,000 Twitter followers that year by posting open-source projects. Free access is harder because model inference still costs money.
Hassan describes two ways to approach infrastructure credits:
- With an audience: Pitch a planned launch and expected usage to a provider, offering attribution in the application footer and README in exchange for credits.
- Without an audience: Publish a high-quality open-source project with a spending limit—he suggests $50—then use actual traction and the provider’s visible role in the project to support a credit request.
The second route gives a potential sponsor something concrete to evaluate. It does not guarantee support, but Hassan says he has seen it work for people without an existing following. He names Replicate, Bytescale and Neon among the sponsors helping keep his own applications free.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Spend time on the interface—and give yourself time to improve
The small model integration does not mean the whole product takes little care. Hassan treats a good-looking interface as essential and estimates spending about 80% of his project time on UI. As a non-designer, he starts by looking at several existing websites—roughly five—and combines ideas he can recognize as effective. The model call may be the novel capability, but the interface is how users encounter it.
Use the stack that lets you work effectively. Hassan’s own choices are Next.js, TypeScript and Tailwind, with Vercel for deployment. Familiar tools let him spend the limited project time on the application rather than learning an unrelated infrastructure stack.
His schedule is a sprint cadence, not continuous weekend work. He describes roughly one focused building weekend per month, sometimes working about 12 hours on Saturday and 12 on Sunday, followed by two or three weekends spent relaxing, watching Netflix and seeing friends. That is a personal arrangement, not a prerequisite: marriage, children or other responsibilities may make a few hours spread across several weekends more realistic.
The final constraint is practice. People sometimes ask Hassan how he builds so quickly because their own projects feel slow; when he asks how many they have built, the answer is often two. His analogy is a second visit to the gym followed by checking for biceps. Improvement requires repeated work over months, and coding is no different.
Hassan says that his everyday job does not involve spending most of his time writing code and that he learned to code only a few years earlier. His closing advice is to find a cadence that fits your life and keep building. Weekend speed is something developed through completed projects, not a secret technique you must discover before starting.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The initial room-redesign application, connecting a Next.js API route to ControlNet on Replicate and image storage on Bytescale.
A face-photo restoration application using GFPGAN through Replicate. Its README flags incomplete setup instructions following an authentication change.
A CLI that generates Git commit messages from staged changes.
Source code for the AI QR-code generator demonstrated in the talk.
The speaker's open-source application for generating Twitter bios with AI.
The original implementation of the face-restoration model used by RestorePhotos.
Further reading
The original paper explaining how spatial conditions such as depth and edges guide pretrained text-to-image diffusion models.
A contemporaneous account of v0's prompt, edit and export workflow for React interfaces.
Full transcript
Loading transcript…