一番下に日本語版もあります

AI Agent Researcher / Software Engineer — Researching how AI agents select and safely execute resources, and running AI services in production.

At Keio University Graduate School of Media and Governance, I research and develop AI agents.

My current focus is less on LLMs themselves and more on the foundational technologies that connect LLMs to real applications.

Alongside research, I build, launch, and operate web services including ChatCore-AI. I develop the backend, frontend, and infrastructure needed to run AI features reliably in production.

• Focus: Object Routing, AI Agent Infrastructure, Multi-Agent Systems, RAG, Browser Agents, UI/UX.
• Location: Japan | Languages: Japanese (Native) / English

I aim to implement AI agents in society to foster a world where human communication flourishes. I want AI to be woven naturally into everyday life — present in the robots and displays we pass simply by walking through the city. By delegating daily tasks to AI, I want to create a society where people can truly focus on meaningful, face-to-face interactions.

I study Object Routing: selecting a set of task-relevant, executable resources from large catalogs of Memory, Skills, Tools, and Agents.

Beyond candidate retrieval, I evaluate the selection pipeline across dependency resolution, conflict handling, permission checks, decisions to abstain or request elevated permissions, and hierarchical routing.

I also study how agents safely execute selected resources, maintain state and context, and recover from failures.

I am implementing these research areas in Marmo-Core, a kernel for routing and safely executing agent resources.

• FIT 2025 (24th Forum on Information Technology, September 4, 2025) — Presented “On Database Structures for Retrieval-Augmented Generation in AI Chatbots” (F-029). I compared document formats for building vector databases from FAQ data. JSON and JSONL achieved the highest success rate in the 100-question evaluation, while CSV performed best in the 1,000-question evaluation, showing that the effective format can depend on dataset scale. (Program)

• NCSP'26 (Honolulu, Hawaii, February 27–March 2, 2026; presentation on February 28) — Gave an oral presentation in English on “Development of Cooperative AI Agents for Improving the Efficiency of Household Tasks”. The system combined an orchestrator with four specialist agents—Browser, Lifestyle/RAG, Scheduler, and IoT—connected through MCP. Across 10 collaborative task scenarios, combining long- and short-term memory improved task completion by approximately 1.7× and made confirmation questions unnecessary. (Program)

My current master's research extends this work to AI-agent resource routing and safe execution infrastructure.

• Resume: https://github.com/kota-kawa/resume
• Portfolio: https://project-kk.com/
• LinkedIn: https://www.linkedin.com/in/kota-kawa/
• Email: kota7kawagoe@gmail.com

Direct access to my currently active applications and platforms.

• ChatCore-AI — Production AI chat platform with multi-model streaming and web tools.
• FS-QR — File Sharing & QR Code Generator.

1. ChatCore-AI (Live) — Production AI chat platform.

   • Overview: Built with FastAPI and Next.js, this live platform supports streaming responses from multiple LLMs, web research and tool use, Tasks and Skills, persistent context, prompt sharing, and generative UI. It also includes Redis-backed session management, usage and cost controls, SSE recovery, tests, and CI/CD, with ongoing improvements for production operation.
   • Technologies: Python, FastAPI, Next.js, React, TypeScript, PostgreSQL, Redis, Docker, Nginx, GitHub Actions, OpenAI, Anthropic, Groq.

2. Marmo-Core — (Lightweight AI-Agent Kernel)

   • Overview: A lightweight Python kernel for registering, retrieving, selecting, and safely executing AI-agent resources. It provides policy-gated execution, audit trails, human approval, recovery, a CLI, and integrations with OpenAI-compatible and Anthropic models.
   • Technologies: Python, LLMs, retrieval, policy and security, CLI, JSON resource definitions.

3. Browser-Agent — Control the web with natural language.

   • Overview: A browser automation agent with a FastAPI web interface. Executes tasks via natural language instructions with real-time monitoring via noVNC.
   • Performance: WebArena Shopping task success rate: 32.6%.
   • Technologies: Python, FastAPI, TypeScript, Docker.

• Keio University, Graduate School of Media and Governance
  • Master's Program, Cyber Informatics (CI) Program
  • Kanagawa, Japan | April 2026 – Present
• Kanagawa Institute of Technology (KAIT)
  • Faculty of Information Technology
  • Department of Information Network and Communication
• University of North Alabama — Exchange Program Completed
  • Alabama, USA | May 2023 – May 2024

• Researching / Building: I am developing Marmo-Core, focusing on its resource selection pipeline and execution controls.

Last updated: 2026-09-23