Profile
Education
- 2019 – 2022Ritsumeikan Keisho Senior High School
- 2022 – 2026Ritsumeikan University, College of Information Science and Engineering
- 2026 –Ritsumeikan University, Graduate School of Information Science and Engineering
About Me
Originally from Hokkaido, I played rugby throughout high school. At university, I conduct research in the field of service computing. Outside of academics, I work as a long-term intern at an AI development company.
In my free time, I enjoy working out, traveling, and exploring good food.
Research
Background
In recent years, composite services that combine multiple Web services (e.g., travel booking sites that offer flights, hotels, and maps in one place) have become widely used. However, selecting the right services from a vast number of options and combining them appropriately is extremely labor-intensive. Traditional approaches required users to describe their requirements using formal logic expressions, making it difficult for general users. Although the emergence of LLMs has addressed this challenge, recommending from a vast number of APIs with a single LLM leads to degraded inference accuracy due to context length limitations and prompt bloat. This research proposes a multi-agent cooperative system to solve this problem.
Research Details
This research proposes a system that leverages Large Language Models (LLMs) to automatically recommend appropriate Web service combinations based on user requirements written in natural language.
Delegating everything to a single LLM causes accuracy degradation due to excessive information volume. Therefore, we constructed a multi-agent system where multiple LLM agents collaborate through role-based division of labor.
We adopted the Contract Net Protocol (CNP) for inter-agent coordination. In this framework, a "Manager" issues tasks and "Contractors" bid on them based on their specialized domains. The Manager publishes user requirements, each Contractor proposes based on their expertise, and the Manager ultimately determines the optimal combination.
After comparing three role-division approaches, the Manager-driven approach achieved the highest accuracy, outperforming the single-agent approach.
Work
ShigaChat
Team Project In use at Shiga International Association
Shiga Chat is a multilingual, restricted-access Q&A service designed for staff at the Shiga International Association.
Developed as part of our research lab's activities. We continued implementation through to deployment, and it is currently in operation at the Shiga International Association.
By combining ChatGPT with Retrieval-Augmented Generation (RAG), it provides fast, region-specific answers to questions about daily life. For each user query, RAG searches and references an existing Q&A database, passing the relevant text to ChatGPT, which then generates a natural-language response.
Key Focus Areas
- User Experience: Designed an intuitive flow from question submission to answer delivery. Prioritized screen transitions and operability to ensure usability for first-time foreign users.
- Accuracy & Safety: Built a curated Q&A database to prevent ChatGPT hallucinations. Introduced manual content review and multilingual grammar checking.
- Multilingual Support: Designed all core operations including notifications and search to be fully available in all supported languages, with cultural sensitivity in mind.
Reference Q&A Data
Shiga International Association Daily Life Consultation Q&A
Diary Board
Team Project
Diary Board is a multilingual educational support tool designed to prevent involuntary immigrant children from becoming isolated in Japanese schools.
This system was developed as part of our research lab's activities, through visits to the Shiga International Association where we identified their challenges through interviews.
Through diary entries, students learn about the multicultural lives of foreign children, breaking down barriers caused by cultural and linguistic differences and fostering inclusive classroom communities.
Key Focus Areas
- Encouraging Continued Use: Implemented ranking and achievement badge features to motivate students to keep writing diary entries.
- Multilingual Learning Support: Added quiz features based on diary content, enabling students to learn languages in an engaging way.
Long-Term Internship
InternshipRoles
End-to-end ownership of backend, frontend, and infrastructure design and implementation.
Created and presented live demos during client meetings to directly convey technical value.
Conducted technical screening and interviews for engineering intern candidates.
Development Achievements
Full Automation of Office Tasks via AI
Combined AI agents with RAG, OCR, and Embedding to fully automate clerical tasks such as invoice processing and DB registration, handling notation variations and irregularities like a human.
AI-Powered Internal Search System
AI searches internal documents across departments and responds to natural language queries. Automated knowledge sharing previously dependent on senior employees, significantly reducing search time.
AI × OCR Document Digitization
AI analyzes OCR-extracted text with contextual understanding, handling notation variations and irregular formats. Automates the entire pipeline from document processing to database registration.
Demand Forecasting
Predicted and analyzed historical sales data using machine learning models to automatically calculate shipment volumes.