Image Annotation AI
Introduction
Designing For Scale With Machine Learning
In 2020 I worked with a startup called Image Annotation AI. They were at that time selling image annotation services, datasets, and machine learning services to companies with specialized ML needs. They had created a distributed workforce that would capture original photography of select objects and annotate them. Then they would also train models according to spec and return these to the client. Clients had specialized needs like in agriculture, medicine, retail and so on. Just to give you an example, Walmart may have a need for retail shelf image annotation, in order to train ML models that can detect the products on their shelves through their security cameras.
I joined the company to work with Andy Gough who I had worked with in the past on his Microwork company and helped him set up his distributed workforce. In this role I worked as Head of Design. My tasks ranged from basic design tasks such as setting up marketing websites, helping design the app for workers, and improving their internal dataset management tool. But for this case study, I will tell you about one project we did at Image Annotation AI: Dataset Studio. This was a large endeavor meant to not only solve a major business bottleneck issue at the company, but also greatly increase the potentail future of the company through an open platform.
I joined the company to work with Andy Gough who I had worked with in the past on his Microwork company and helped him set up his distributed workforce. In this role I worked as Head of Design. My tasks ranged from basic design tasks such as setting up marketing websites, helping design the app for workers, and improving their internal dataset management tool. But for this case study, I will tell you about one project we did at Image Annotation AI: Dataset Studio. This was a large endeavor meant to not only solve a major business bottleneck issue at the company, but also greatly increase the potentail future of the company through an open platform.
The Problem
B2B Sales Flow Lacked Speed And Scalability
The client workflow at Image Annotation AI was to conduct business calls with client representatives, intake their requirements, create tickets for the workforce, and then conduct a series of calls to check in and hand off the deliverables. We hypothesized that through an open self service platform, clients would be able to more quickly move through these steps with or without a sales rep, and have more direct visibility and control over their dataset through the process and handoff.
We observed that during peak times of demand, our staff had trouble scaling to meet the demands of clients, but during down times, they lacked productive and meaningful tasks to take on. So we thought that a self service platform where clients can create requests in the app itself, which automatically distributes the tasks to our workforce, and then updates them with the results, would be a faster and easier process that would scale much more easily. This new solution concept we called Dataset Studio.
We observed that during peak times of demand, our staff had trouble scaling to meet the demands of clients, but during down times, they lacked productive and meaningful tasks to take on. So we thought that a self service platform where clients can create requests in the app itself, which automatically distributes the tasks to our workforce, and then updates them with the results, would be a faster and easier process that would scale much more easily. This new solution concept we called Dataset Studio.
The Process
Stakeholder Interviews, User & Market Research
Through working with the company, observing client calls, and interviewing staff, I came to understand their working process. And talking with Andy, we came up with this idea together of an open platform that would significantly grow the company through scalability, and potential future features such as community features, marketplace, and so on.
I designed Dataset Studio through a user centered design process that began with interviewing relevant staff members and market research. Then I sketched solutions and discussed them with Andy and a couple product team members. After this I wireframed the concept and continued iterating and gathering feedback until the visual design was shaping up. I used Material Design to start as a base for our own design system, and altered the styles and colors to have something custom for our new platform.
I designed Dataset Studio through a user centered design process that began with interviewing relevant staff members and market research. Then I sketched solutions and discussed them with Andy and a couple product team members. After this I wireframed the concept and continued iterating and gathering feedback until the visual design was shaping up. I used Material Design to start as a base for our own design system, and altered the styles and colors to have something custom for our new platform.
The Solution
Our Solution, Dataset Studio, Featured Three Main Ideas.
1) Anyone can use the platform by signing up and creating a profile. Use for them would start free. They would be able to do their own dataset creation, image annotation, and machine learning model training. We hypothesized that this would grow our user base and a percentage would pay for premium services like original image capture and annotation.
2) Clients would be redefined from just the large enterprise clients to other business sizes as well, and they would submit their orders through the platform. We would still have our old workflow of meeting with clients one on one if that is what they prefer. But the new platform would be an option that we would hope would meet more clients needs faster. Through the platform they can submit orders for image datasets, annotation, and model training.
3) We imagined that in the future this platform would also enable community features such as sharing or selling datasets and models or social discussion features. By creating an open version of our platform we could potentially scale into these ideas later on.
2) Clients would be redefined from just the large enterprise clients to other business sizes as well, and they would submit their orders through the platform. We would still have our old workflow of meeting with clients one on one if that is what they prefer. But the new platform would be an option that we would hope would meet more clients needs faster. Through the platform they can submit orders for image datasets, annotation, and model training.
3) We imagined that in the future this platform would also enable community features such as sharing or selling datasets and models or social discussion features. By creating an open version of our platform we could potentially scale into these ideas later on.
Impact
Proven Concepts Now Popular
Though Image Annotation AI shut down, Dataset Studio created a lot of impact by planning for an expansive and successful future for the company. We conducted user and market research and made a plan that was based on sound data and predicted future trends. As evidence, you can look at social AI platforms like Civit AI and Hugging Face, which are using some of the concepts we discussed years ago. Sadly due to a lack of sufficient funding and investment strategy they did not have the funds to pull off building Dataset Studio so it never got built. But by looking at other companies that are doing something similar to what we imagined, in the realm of image annotation look at Scale AI, we can confidently say this would have been a successful project if it had sufficient funding behind it.
