Mr. Ryo Takashina, Representative Director of APTO, Inc.
Developing AI requires a huge amount of data points―as many as tens to hundreds of thousands, or even more than that. In addition to quantity, the quality of data also affects the accuracy of AI. APTO, Inc. offers harBest, an AI development platform that can collect large amounts of high-quality data quickly and at a low cost. This is an interview with Mr. Ryo Takashina, the Representative Director of the company.
―Could you tell us about APTO, Inc.?
Representative Director of APTO, Inc.
Founded in 2020, APTO offers an AI development platform called harBest. It is extremely laborious and time-consuming to collect, prepare, and annotate data during the development of AI. The primary feature of harBest is to enable customers to conduct these processes at a high quality and for a low cost.
―What inspired you to start the business?
After graduating from college, I worked as an engineer developing major core systems. I also worked as a systems engineer for an IT consulting firm before founding a VR-based content production company in 2017. At that company, I was trying to develop a service that would automatically detect flames on social media, but the accuracy of the service didn’t quite improve. I investigated why the service wasn’t more accurate and found that the causes were insufficient data and poor data accuracy. At that time, however, there were no services yet that could easily collect large amounts of good quality data. So, I decided to found APTO.
―Where does harBest’s name come from?
The name comes from “harvest.” It means companies can harvest data and crowd workers (hereinafter simply called “workers”) can harvest data points.
The name “harBest” contains “B” instead of the “v” in “harvest” to indicate that companies can collect the best possible data and that workers can earn points in the best possible way.
―What kind of service is harBest?
harBest is a service for resolving issues in the data preparation and annotation processes of AI development.
Conventionally, this work process has been cumbersome, requiring many redundant interactions and a lot of time and effort. With harBest, however, companies can manage the process centrally and reduce the work time and costs by more than half.
AI has already been introduced into various products, but do you know how AI is developed?
There are four major processes involved in developing AI: collecting and preparing data, building the AI, incorporating it into a system, and then evaluating the results. Based on the evaluation, data will be collected and prepared again. This is the typical timeline of AI development.
The most time-consuming process among them is the data collection and preparation. It’s said that this process accounts for about 80% of the development time. In the AI industry, this data preparation is called “annotation.”
Annotation is the work of assigning correct answers. For example, in order to teach AI dogs, it is necessary to specify which part of an image of a dog against a background is the dog. In addition, tens of thousands of images in which the boundaries are precisely specified may be needed to make AI recognize dogs
Collecting such a large amount of data isn’t so easy. This is where harBest comes in.
On the harBest management screen, a company user selects their desired task type and enters the project name, budget, and worker attributes. If necessary, they perform additional operations, such as uploading images. Pressing the start button at the end sends the task to our registered workers. The workers can then earn points for just doing simple work. This system allows companies to complete large tasks quickly, inexpensively, and accurately through crowdsourcing from our workers.
Here, I’d like to mention the importance of data. A report about the task of improving an algorithm says that the accuracy of the algorithm is far more enhanced by improving the data than by improving the algorithm itself. This means that high-quality data is one factor that increases the probability of success of AI projects.
So, we decided to collect reliable, high-quality data on harBest using an automatic check algorithm based on work done by multiple people.
In addition to this algorithm, we have ensured that a smart checker AI checks and rates workers according to the accuracy of their work in order to increase the reliability of the data. Since the number of points workers can earn changes according to their ratings, this gives workers incentives to produce good quality data.
The work management screen also allows companies to view data in a list so that they can easily perform operations such as rejecting or removing poor quality data, if any.
Another feature of harBest is its professional user system. A professional user is a worker who has been authenticated, signed an NDA and a pledge, and passed an annotation test.
Some requests from companies may involve confidential data. In such cases, companies can order services from professional users.
―What cost advantages do you offer?
Since our platform allows companies to order directly from workers, there are no brokerage costs, and we can offer dramatically lower prices compared to our competitors.
Let’s take image classification work, for example. A company can do it in three days for 50,000 yen with us compared to two weeks for 200,000 yen with our typical competitors. The reason why we can offer such quick and low-cost services is because we have hundreds of active users (workers). In addition, 63% of our users are weekly active users, which is a very high activity rate.
Tasks that are going on right now on our platform include detecting wheel objects and finding illegal posts. The former is the task of framing wheel parts in photos. It will take only two or three days to complete a deal for 6,000 items. The latter is the task of checking to see if the text displayed is an illegal post. To develop AI that checks posts for illegality, the detection work is usually done by three people and then a majority vote is taken. It will take only three or four days to complete the detection work on data of 15,000 posts.
Companies can also submit other tasks, such as paint-over work and photo-taking work.
We have received positive feedback from companies saying that they have been able to reduce their costs as a result. In addition, we have heard from workers that they have been able to earn more than with other apps.
In the year since its release, harBest has processed 2 million pieces of data. We are confident that our platform has processed a significantly larger amount of data than other companies’ services do.
AI development is expected to continue to be active in the future. harBest enables companies to develop AI more efficiently. We hope that those interested will start with our free trial.
| Company name：APTO, Inc. |
| Founded：January 2020 |
| Main businesses：Providing an AI development platform and consultation on AI URL：https://apto.co.jp |
This article is part of a series of articles introducing venture companies working together as ICF members to resolve societal issues.