Dataset Studio

Original Design System for an AI ML Product

Dataset Studio was an AI ML creation platform based on image annotation and dataset creation workflows at a company called Image Annotation AI. In 2020 I had worked with the founder, Andy Gough, on product and marketing for his company. At the time we were promoting the image annotation services of his existing business workflow. Through conversations, he told me he had been invisioning a platform where his internal process would become open to the public. The process they had until then was to intake a business with machine learning dataset needs, then based on what they needed, distribute the work of image capture and annotation to a distributed workforce he had created.

I had worked with him in 2015 on creating an app called Microwork IO that was about distributing these tasks via mobile phone to allow people across the globe to earn money capturing photos of everyday objects. The value in doing this is to create a kind of dataset in machine learning that is called “in the wild” data. This kind of photographic data has value in the ML world because you can create custom and specific datasets around particular problems and environments with a lot richer data to train your models on. If you have particular ML needs, getting the right kind of dataset through a premade dataset you can find online can be difficult or impossible. Andy had identified this niche problem in the ML industry and decided to address it early on in 2015 with his image annotation business using his distributed workforce, utilizing the mobile app I had strategized and designed for him.

In 2020, we revisited the strategic direction of his business. In 2020 he had been working as a B2B business operation where businesses like Wal Mart and others would hire him on a bidding basis against other image annotation and dataset providers. They’d compete on issues like features, delivery speed, price, data quality, human annotation features, etc. Based on whether or not Andy could win a contract, his business would succeed or fail. This was the state of his business when I met with him to begin work together once again in 2020.

In response to the state of his business and the challenges he was facing, we decided to take a new direction that would fundamentally open up the platform and drastically increase the available opportunities for success. He had identified that his business was basically competing for projects, and we were planning a pivot that would make his business more independent and able to grow on its own without constantly attaining new deals. So, to this end, we created Dataset Studio.

What Dataset Studio would do is allow anyone with an account to create their own datasets in a web application. They could upload images, annotate them, and train an AI model all inside the app. And Dataset Studio would address pricing later on through subscriptions and premium services like distributed original data collection (the Microwork model mentioned earlier), annotation (the current Image Annotation AI services), or possibly more. These would be offered as in-app microtransactions during the user’s ML creation flows. For instance, while annotating images, there could be a call to action to automate this process with the human distributed workforce instead for a price, and a purchase flow to create that order.

In 2020 I had completed this work for Andy that included a full product design utilizing Material Design as a design system bootstrap, and an original brand identitiy for his various business surfaces that would tie them all together to one single brand voice. Many businesses, but one brand voice. I also had complete prototypes and feature planning for creating datasets and annotating them in the web app.

In 2024 I was looking for a project for my portfolio to focus on design systems. I decided to revisit Dataset Studio even though Andy had stopped running his businesses some years ago to retire. So in 2024 I looked at Dataset Studio and decided to use it for this design system project. I would first redesign the app at the high gestural level to create an original visual language, then I would create a simple design system based off the designs I had made, and finish with a design token system and Storybook library with CSS tokens. I finished by using VS Code, HTML, CSS, JavaScript, and ChatGPT to create the Storybook app. It ran aspects of the design system in a browser window where you could mouse over certain colors or fonts and then the color or font would turn into a CSS token that you could click on, and doing so would copy that code to your clipboard. In this way, developers could look at the styles in the library which would act as a source of truth and easily copy and paste the styles. On my end, I could change the colors or properties of these styles without changing the CSS. So that when a developer inputs for example background-surface-0, that code stays static and doesn’t need to update, but if I need to change the color of that style I could change it in Storybook and it would update in the developer’s CSS file without changing every place that token, background-surface-0 is referenced.

The point of this exercise was to create a modern AI app with a design system for my design portfolio. I had taken a class on design systems and wanted a portfolio piece specifically dedicated to this recently popular topic. By doing so, I was able to update the concept of Dataset Studio and demonstrate a modern design system in Figma and Storybook. The new concept of Dataset Studio would utilize community features that are more like the modern AI landscape when looking at apps like Civit AI and Hugging Face. And the design system, instead of essentially being Material Design before, is now an original design system with original styles and brand expression concepts. This project was a great learning experience and chance to express many technical and artistic ideas at a high gestural level.