Hello everyone! This is Shu from the Global Operations Team (GOT).

As the use of AI accelerates across the company, we are at a pivotal moment in how we work. While there’s excitement about the potential for AI to make our daily tasks more efficient, some may worry about how their specialized skills will remain relevant in the future.

As specialists in interpretation and translation, GOT decided to tackle this question head-on. We’ve come to believe that AI is not a replacement for our expertise, but rather a powerful tool to enhance its value and expand the scope of our work.

In this article, we’ll share the full story of the “GOT AI Hackathon,” our first concrete step on this journey. We hope that documenting our initiative will serve as a helpful reference as your own teams explore how to leverage AI.

Why a Hackathon? Fostering a Culture of Learning by Doing

To kick off our AI utilization, we chose to host a hackathon rather than a traditional training session or seminar. We believe that in the age of AI, it’s more important to gain practical knowledge through hands-on experimentation and trial and error than to simply absorb information.

The goal of this hackathon wasn’t to build a perfect, finished product. Instead, we approached our initiative with three aims:

1. Get comfortable and familiar with AI.
2. Discover new possibilities for our work by interacting with AI.
3. Share our processes and learnings as a team to strengthen the entire organization.

In a typical hackathon, participants might dedicate all their time to development. However, Mercari’s interpretation and translation support is a critical operation that cannot be paused for even a single day. Therefore, our hackathon was structured as a two-week challenge where members juggled their regular duties while working on their projects.

Concrete Initiatives and Results from the Hackathon

The hackathon produced a wide range of ideas, from small improvements to our day-to-day work to ambitious concepts for overhauling entire work processes. Here are some specific examples and the lessons we learned from them.

Company-Wide Tools Born from the Hackathon: Turning Specialized Know-How into an Organizational Asset

A major success of this initiative was seeing hackathon prototypes evolve through further refinement into tools that are now used company-wide.

Meeting Minutes Translation Tool (@mandrill)

• Challenge: In meetings where multiple languages are spoken, creating and sharing minutes in both Japanese and English is essential for supporting transparency as well as inclusion and diversity (I&D) at Mercari. However, preparing these bilingual minutes in real-time was a significant burden.
• Approach: To solve this, @mandrill built a prototype translation tool that runs on Google Docs. Using Apps Script code generated by GPT-4o, she tested the feasibility of a feature that automatically formats and translates text into a side-by-side Japanese-English layout.
• Outcome: After clearing security requirements and incorporating learnings from the hackathon, this tool has been officially released internally. All employees can now use it in Google Docs. This has not only significantly reduced the person-hours required to create minutes but has also allowed members to follow detailed discussions where they previously struggled due to language barriers. 

Bilingual Slide Creation Tool (@mandrill, @celeste)

• Challenge: Creating bilingual slides, especially for company-wide meetings (All Hands), while maintaining a clean layout was an extremely time-consuming task.
• Approach: @mandrill and @celeste tackled this by developing a tool that generates translations next to or below the original text, making it easy to display both Japanese and English on a single slide. They built the prototype using a combination of technologies, including Apps Script, and “Ellie ” (GPT-4), and Gemini 2.5 Pro.
• Outcome: After the hackathon, the tool was further improved, particularly in its ability to maintain the original layout. This feature is now a standard part of the official Google Slides template for All Hands meetings. Now, presenters can create their content and then select the function from a menu to generate a bilingual version with minimal effort.

Spreadsheet Translation with Internal Terminology (@celeste)

• Challenge: Translating spreadsheets efficiently while ensuring the accurate use of Mercari’s internal terminology was a complex problem.
• Approach: @celeste took on this challenge using Apps Script and Gemini. For security and cost reasons, she opted against using external translation APIs and instead implemented a new LLM proxy service developed by our internal security team.
• Outcome: Although directly leveraging the termbase in the script proved technically difficult at first, improvements continued after the hackathon. The tool is now available company-wide, allowing anyone to translate a spreadsheet while reflecting the internal termbase with the click of a button.

---

The creation of these three tools was made possible not only by the efforts of GOT, but also thanks to helpful advice on security and accessibility from our colleagues on other teams.

---

Diverse Approaches to Improving Daily Work Processes

Improving Research Efficiency and Quality (@sherry)

• Challenge: Researching official translations for technical terms was time-consuming. In large-scale translation projects involving multiple members, maintaining consistency in terminology was also difficult.
• Approach: @sherry utilized Gemini (2.5 Flash) to address this. She created and tested a prompt that instructs AI to read a Google Docs URL and list specialized terms, their English translations, and the official source URLs that justify the translation.
• Outcome: This approach confirmed that AI can search websites and present translation candidates along with their sources. However, it also highlighted the current limitations of AI, as the number of terms in the output fluctuated and manual adjustments were still necessary.

Supporting Content Creation for Public Communications (@sherry, @emma)

• Challenge: It takes significant effort to craft concise, engaging social media posts for platforms like LinkedIn from long-form articles about Mercari’s initiatives.
• Approach: @sherry and @emma used Gemini (2.5 Flash) to summarize long articles about Mercari and draft LinkedIn posts.
• Outcome: While the AI could summarize articles, the results were often verbose or failed to strike the right tone for a corporate account. We found that a more effective approach for now is to have the AI generate multiple drafts, which a human editor can then combine and refine.

Streamlining Client Communication (@Eri)

• Challenge: The process for handling interpretation requests in Jira was cumbersome. A team member had to manually provide an initial response in a dedicated Slack channel, and matching Jira and Slack usernames was often difficult.
• Approach: @Eri built a “Booking Bot” by combining tools like “Ellie” (GPT-4o), Zapier, the Slack API, and Webhooks. The bot is triggered by a Jira request and automatically posts a preliminary reply in Slack. Specifically, it receives information via a Jira Webhook and replies with a mention to the requester in a Slack thread.
• Outcome: The bot successfully automates initial responses and reminders, notifying requesters with a mention and significantly reducing the manual work involved in client communication.

Continuous Process Improvement Initiated by the Hackathon (@yuko)

• Challenge: The interpretation request workflow was a deep-rooted problem, with information scattered across multiple tools like Slack, Jira, and Google Calendar, leading to high-context switching loads and communication costs.
• Approach: Instead of focusing on a single tool, @yuko took a holistic view and focused on “improving the entire interpretation request workflow.”
• Outcome: During the hackathon, she created a prototype for a workflow that could be completed entirely within Slack, bypassing Jira. Based on this idea, the team continues to discuss how to build the ideal workflow. The “just try it” momentum from the hackathon led to another concrete success story shortly after: @yuko, with guidance from AI, quickly developed and released a GAS script that automatically generates announcement slides with embedded code for listening to interpretation on Interactio. This is a prime example of how learnings from the hackathon directly led to the development of practical tools, showing that the momentum for AI utilization continues within the team.

Efficiently Formatting and Aligning Translation Assets with AI (@Miguel_RM)

• Challenge: Previously translated files are valuable assets, but many are saved in formats that are difficult to reuse in a translation memory (TM). We also faced inconsistencies between new machine translated (MT) content and past human translations as the use of MT became more widespread.
• Approach: @Miguel_RM tested the idea of using generative AI (Gemini) to accurately align old source and target files sentence by sentence, formatting them into a unified, TM-friendly format. The goal is to efficiently incorporate past translation assets, including post-edited MT output, into our TM to manage consistency.
• Outcome: This approach showed the potential to drastically streamline the alignment process, which would be incredibly time-consuming if done manually. At the same time, it revealed the technical limitations of current AI in processing large, complex datasets at once. It was also a good lesson in the trial-and-error nature of prompt engineering.

Automating Interpreter Assignments with a Focus on Well-being (@Shu)

• Challenge: As an interpretation coordinator, I find the work of assigning GOT interpreters to be a complex, puzzle-like task. It involves optimally allocating a limited number of interpreters to a constantly changing slate of requests, all while adhering to detailed team rules about working hours and assignment methods. This consumed a significant amount of the coordinator’s time.
• Approach: I articulated this complex decision-making logic into a detailed “rulebook” and then implemented it as an automated script using AI and GAS.
• Outcome: The script makes decisions based on rules designed to ensure quality, such as minimum staffing requirements and strategic resource allocation based on meeting length. It also incorporates rules that protect interpreter well-being, such as limits on consecutive assignments and ensuring break times. This automation has led to a significant reduction in person-hours and the risk of human error.

Beyond Efficiency: GOT’s Vision for the Future of Communication

This hackathon confirmed for us that AI is a powerful tool for improving operational efficiency. We intend to reinvest the time and resources saved through our collaboration with AI into creating more fundamental value. This could mean building a communication infrastructure that more powerfully promotes Mercari’s I&D. Alternatively, it might involve providing more advanced communication support that bridges not only language barriers but also differences in cultural backgrounds and contexts.

Moving forward, GOT will continue to embrace AI as a unified team. Our goal is a future where we not only improve our daily work but also deliver value that was once only possible for human experts, deploying it across the company through new tools and systems.

We hope this article serves as a helpful reference as you and your team consider how to engage with the new tool that is AI.