Mercari aims to become one of the world’s leading companies in the field of AI. This vision is not simply about the pursuit of technological reform but is rather an expression of our strategic efforts to transform the very core of our user experience. The team playing a central role in achieving this vision is the AI/LLM Team (also known within the company as the Eliza Team).

This team works not only to pioneer practical usage of new technology such as generative AI and LLMs but also to leverage our existing assets in AI to make our service as a whole more intelligent. Within about a year of the team’s establishment, it has achieved tangible results, including major click rate improvements and the groundbreaking implementation of our AI Listing Support feature.

But this is only the beginning for them. From building the foundations that will allow us to leverage AI company-wide to creating innovations in user experience, the team’s sights are always set on the future. We spoke with team leaders Max Frenzel (@maxfrenzel) and Ryan Ginstrom (@ryan) to hear the details of the team’s achievements so far and their plans for the future.

According to the two of them, the future they envision will combine technology and user value in a way that is deeply unique to Mercari.

Building and enabling—the two pillars of Mercari’s AI strategy

── Could you start by telling us about the background behind the establishment of the AI/LLM Team?

@Max：Our team was established to serve as an important turning point in Mercari’s AI strategy. Our team has two important roles: “building” and “enabling.” Building involves leveraging generative AI to provide new value and a new experience to our end users. The development of AI Listing Support, a feature that leverages AI to support users’ daily actions, falls under this category.

Enabling involves enhancing the productivity of our employees and the company as a whole. By improving and automating work processes to achieve more efficient ways of working, we help to improve the productivity of the entire organization.

── What is the current progress in each of those categories?

@Max：We are currently looking to be a driver of change with AI on a broader scale, testing our AI initiatives and collaborating with other feature and product development teams to create greater value.

The key here is adapting the AI capabilities of individual teams into a form that can be used company-wide. For example, we are working to achieve things like making image recognition technology developed by one team usable by other teams, and creating a platform for sharing natural language processing features throughout the entire company. This will increase the reusability of the technology and improve development efficiency.

── Mercari has a long history of leveraging AI and machine learning. How is your team positioned within all this?

@Max：Mercari is strong in AI and has long utilized ML. However, until now each team’s efforts have tended to be isolated, and there has not been sufficient collaboration or sharing of resources and features developed across teams. We hope to change this by bringing together the various existing initiatives around AI and creating one unified effort as a company.

In particular, we are positive that combining our past AI initiatives with today’s generative AI and LLMs will bring about transformative results. For instance, combining our existing image recognition technology with generative AI will allow us to build more sophisticated item recommendation systems.

@Ryan：The AI department used to be a single unit, but now our members are distributed amongst various teams throughout the company. This change allows us to develop more closely with our customers, and allows teams to develop based on more specific use cases. At the same time, we are focusing on maximizing the strengths of the company as a whole by making sure that the technology and knowledge are shared.

Aiming for a world-class technology infrastructure

── Could you both tell us more about your roles and responsibilities?

@Max：I joined the team in November of 2023 as a product manager, and since March or April of 2024 I have been leading the team as a whole, overseeing both the product and engineering sides. My main roles are stakeholder management in the collaboration between feature development teams, identifying user pain points, and considering solutions. I am also responsible for formulating the team’s overall strategy and vision, as well as communicating with senior leaders and executives.

@Ryan：I primarily support the team from an engineering perspective. In addition to selecting technologies and deciding how to implement policies from a technical perspective, I also focus on promoting technical collaboration with other teams. I place particular emphasis on building a reusable technology infrastructure that is not confined to the development of individual features. This type of infrastructure improves not only the quality of the product but the development efficiency of the entire team.

── What is the current team structure like, and what are the team’s plans for the future?

@Max：Currently, the team consists of mainly ML and backend engineers. Our members are relatively young, highly curious, and eager to explore new technologies. Our plan is to hire more members to diversify the range of expertise on the team and double its size over the course of the next year.

We are particularly focused on establishing a small-scale research team of AI/ML researchers responsible for exploring opportunities for innovation in the medium- to long-term. This team will focus not only on research but on investigating global AI trends, and will facilitate implementing and leveraging these findings in our actual products.

We also plan to hire client engineers, UX specialists, and data scientists to form a team structure that enables streamlined iteration from ideation to implementation and validation.

The fruits of collaboration

── What are some of the specific results the team has achieved over the past year?

@Max：One of the most impressive results we achieved was through a collaborative project with the Search Team. Our team was responsible for ML research, developing and testing a new model, and implementing it in the production environment, which resulted in a 13% improvement in click rate of similar items. This could not have been achieved without close collaboration with the Search Team.

We were able to achieve massive improvements by effectively leveraging the existing ML infrastructure and through rapid iteration and testing. This project also helped create a model case for cross-team collaboration.

Another accomplishment was the development of the AI Listing Support feature, for which we took complete ownership of backend development. We were able to achieve major improvements in user experience by increasing the sophistication of image recognition technology, improving the system for auto-tag generation, and enhancing the recommendation feature. Working closely with the UX Team to successfully implement an easy-to-use interface without compromising on technology was definitely a major feat.

── What is your team’s relationship with the Corporate Engineering Team (CET)?

@Max：Currently, the boundary between what each team does is quite fluid. We work with CET mainly to improve productivity throughout the company, and we are currently collaborating on efforts to improve ML systems and reform knowledge management. Our teams have a complementary relationship in which we focus mainly on developing new use cases for generative AI/LLM technology and CET focuses on developing tools that leverage traditional ML technology.

Moving forward, we aim to deepen our collaboration and strengthen our team’s role as an enabler within the company. Specifically, we believe that combining the knowledge CET possesses with the new technology our team handles to develop and implement new AI tools will lead to more powerful solutions.

Accelerating innovations in AI

── What sort of challenges do we face on the technical side, and what approaches are we taking to solve them?

@Ryan：The main technical challenges we’re facing right now are around efficient operation of LLMs, optimization of real-time inference systems, and data quality control and privacy protection. We are taking a phased, systematic approach to tackling these challenges.

In the first phase, we are working to lighten and optimize our models, leveraging edge computing where necessary. At the same time, we are working to strengthen data governance and establish a security framework. These efforts not only improve our technology but are important to maintaining the trust of our users.

@Max：In addition to these initiatives on the technical side, we are also focused on tackling organizational challenges. The key here is to strike a balance between vertical and horizontal approaches. Our vertical approach aims to create direct user value by developing specific features end-to-end. Our horizontal approach involves developing AI platforms and reusable components to establish a company-wide technology infrastructure.

The future value of experience made possible by technology

── Could you tell us more about the new research team you’ll be hiring?

@Max：The research team will be responsible for two key areas: UX research and AI research. For UX research, we plan to bring on board senior talent well versed in generative AI. In the realm of generative AI, users and systems interact in non-traditional ways, so it is particularly important to have a deep understanding of these interactions in order to design optimal experiences.

We plan to continuously improve user experience by validating and prototyping new interfaces based on a thorough understanding of user behavior and conducting user tests. This insight can then be leveraged from the very early stages of product development.

For AI research, we will structure the team around ML researchers and explore the latest technologies. The key here will be to focus not only on research but on the speed at which we can apply research results to our product. The team will regularly develop and validate prototypes and beta test experimental features to achieve a balance of both theory and practice.

As I mentioned earlier, the speed at which we are able to implement research results into our actual product is key. We hope to accelerate the practical application of our research through continuously gathering and analyzing user feedback and close collaboration with the development team.

── To end, please share with us your outlooks on the future.

@Ryan：What I would like to convey to society in particular is the potential that LLMs hold. Today, people still have a rather vague perception of LLMs, but this technology is more than just a demonstration of what could be. I want to make it widely known that this technology has the potential to create tangible value in business and daily life applications.

Within the company, I’d like to continue promoting the use of new technologies. I am certain that by combining the expertise of each team with AI technology, we can create unprecedented value. I also want to create an environment that ensures the passion and ideas of young engineers can take shape in the form of actual products.

@Max：What we’re aiming for is a world in which AI is naturally integrated into the services we use. Take item recommendation, for example: with AI, we can provide more personalized suggestions and user interactions with a deeper understanding of context. AI technology will also enable advanced automation in areas such as quality control and fraud detection.

I would like to encourage us as an organization to leverage AI company-wide and proactively share knowledge across teams. It’s important to simultaneously work toward achieving a unified technology infrastructure while also fostering a culture of continuous innovation.

Moving forward, our focus will be on implementing these changes in a way that feels natural and adds value for our users. Our goal is not the use of AI itself, but the improvement of the user experience. We aim to deliver tangible value in forms such as an easier listing process, more secure transactions, smoother communication, and more.