Please provide a brief self-introduction and your background.

I first encountered machine learning in 2012 when I entered graduate school. At that time, AI was still a rapidly evolving field, and most of my initial work was rooted in academic research. After graduation, I joined a web search engine company, where I witnessed firsthand the explosive growth of machine learning in both research and industry. Companies were beginning to adopt AI not only for experimental projects but also to directly improve the performance of product features, online sales, and advertising delivery.

In 2020, I took on a new challenge by joining a Japanese news aggregator startup. The company was actively exploring the adoption of state-of-the-art machine learning methods to better understand users’ interests. This wasn’t just about recommending news articles — it was about accurately matching content and advertisements to individual user preferences in real time. My primary responsibility was to improve the recommendation model’s performance and streamline the supporting pipeline, a role that required both technical depth and a keen awareness of user behavior.

What aspects of JAPAN AI resonated with you and led you to join the company?

With the emergence of large language models (LLMs) and the rapidly changing ecosystem surrounding them, I became more convinced than ever of AI’s transformative potential. When I learned about JAPAN AI’s clear and ambitious goal of becoming the top AI agent company in Japan, it immediately resonated with me. The determination of the team to pursue that vision — coupled with their lightning-fast iteration cycles — gave me confidence that this was the kind of environment where real breakthroughs happen.

Another strong appeal was JAPAN AI’s forward-thinking approach to AI adoption. The culture here encourages everyone to explore any AI-powered tools that might improve productivity or creativity. We even have an internal target: by the end of the year, at least 80% of our code should be generated by AI. This kind of bold, measurable commitment to embracing AI is rare and inspiring.

Were there any particularly memorable tasks or projects since joining the company?

I joined JAPAN AI just two months ago, so I’m still adjusting to the fast pace and catching up with the latest developments in AI Agent technology. Even so, I was able to contribute meaningfully in my very first month by implementing a smart tool selection mechanism for our agent. This enhancement enables the agent to decide, in real time, which internal or external tools to use for a given task. While it may seem like a small improvement, it has already reduced the operational cost of user interactions with LLMs.

What made this project memorable wasn’t just the outcome — it was the challenge of integrating the mechanism without disrupting the existing pipeline, and ensuring that the selection process remained fast enough to maintain a smooth user experience. Looking ahead, I’m excited about our current initiatives: adding planner mode and memory management so that our agent can hold ongoing, infinite conversations without losing key information.

Please tell me about your team’s mission and your role.

Our team is responsible for developing and maintaining the core agent service, which is essentially the brain and nervous system of JAPAN AI Chat. It links our application with LLMs, intelligently manages and composes the conversation context so that LLMs can fully understand both the environment and the user’s request, orchestrates the steps planned by the LLM, and finally delivers a coherent response.

I serve as the tech lead, which means I’m not only coding but also making architectural decisions and setting technical priorities. Our team is currently just two people — my manager and me — but because our work underpins the entire user experience, the impact we have is significant. We’re also hiring, so expanding the team is a priority.

What are the most technically challenging parts of your work, and how do you tackle them?

Many of the challenges stem from the inherent limitations of LLMs. One persistent issue is the token length limit, which constrains how much context the model can retain in a conversation. Without intervention, this means conversations must be short to preserve full context. We’re addressing this by developing a “context condensing component” that intelligently summarizes and compresses past exchanges while retaining the essential details needed for continuity.

Another challenge is that LLMs are stateless and prone to forgetting important information or generating hallucinations — responses that sound convincing but are factually incorrect. To counter this, we’re building a robust memory management system to maintain context consistency over long interactions.

Cost is another factor that can’t be ignored. Each interaction with an LLM incurs expense, and in a competitive market, cost efficiency is vital. Our smart tool selection mechanism helps minimize costs by ensuring the agent only uses the tools necessary for each task, even as the number of available tools grows. Finally, as a B2B company, we must protect user privacy at all times. This means carefully balancing our drive for more intelligent, capable agents with the need to safeguard sensitive information.

What are the features of JAPAN AI’s product development?

As a startup, our product development cycles are extremely fast. We ship, learn, and improve in short iterations. However, as we gain more customers, we have to be mindful of stability in production environments. It’s a delicate balance — like upgrading a car’s engine while driving at full speed.