​Description

​Join the Tokyo AI community for an in-depth exploration of agentic systems, moving beyond simple prompting into robust, production-grade architectures.

​This session brings together engineers and architects to discuss the fundamental mechanics of long-running inference loops, the implementation of structured "agent skills" for full-stack development, and the strategic design choices required to deploy reliable, domain-specific AI systems.

​This event is designed for technical practitioners looking to bridge the gap between experimental workflows and scalable, maintainable AI applications.

​​Agenda

​We'll start with the foundational layer: understanding the "how" of inference loops and long-running process management sets the technical baseline for the evening. Building on the baseline, we'll move into practical application: how to structure those agents to perform complex, multi-step development tasks. The final talk addresses system-level design and the shift to production, synthesizing the previous two technical talks into a strategic architectural framework.

​​18:00 Doors open

​​18:30 - 19:00 From Inference Loops to Long-Running Agents: Fundamentals, Workflow, and a Real Example (Shekhar Upadhaya, Founder & CTO @ Skyhost)

​​19:00 - 19:30 Agent Skills for end-to-end Full-Stack Development (Kaustubh Hiware, Software Engineer @ Mercari)

​​19:30 - 20:00 Architecting Real-World Agentic Systems (Jeff Yang, Solutions Architect @ Noeon Research)

​​20:00 - 21:00 Networking

​​21:00 Doors close

​Speakers

​Talk 1 - From Inference Loops to Long-Running Agents — Fundamentals, Workflow, and a Real Example

​Speakers: Shekhar Upadhaya (Co-founder & CTO, Skyhost)

​Abstract: Every agent runs on an inference loop. A model, some tools, and a context window. This talk starts with what actually happens under the hood when an agent runs, then gets into a working setup: how I structure repos for agentic work, the tools I use daily, and the patterns that keep long-running tasks from falling apart. We'll close with a walkthrough of a real multi-file refactor that runs for 1-2 hours, what the setup looks like, and what tends to break.

​Bio: Shekhar has spent the last decade building consumer and b2b products as a co-founder, a CTO, and earlier as a software engineer at Amazon. He has shipped across social networking, short video, live audio, and AI companions, reaching hundreds of thousands of users. He is currently building in the live streaming space.

​Talk 2 - Agent Skills for end-to-end Full-Stack Development

​Speakers: Kaustubh Hiware (Software Engineer, Mercari)

​Abstract: Agent skills provide a structured, system-level approach to working with AI, replacing trial-and-error prompting. This talk demonstrates how to use agent skills as building blocks for full-stack development: finding logical loopholes in your implementation plan, defining a unified UX design plan (as a non-designer), and generating a Python backend and React frontend in parallel. We go deeper into designing your own skills, embedding domain knowledge, and creating self-improving memory patterns. We also discuss encoding behavior as a skill versus system context, and how to optimize tokens and observability using statuslines and CLI tooling.

​Bio: Kaustubh Hiware is a Senior Software Engineer at Mercari, building large-scale Search and Product Data infrastructure. He's an experienced speaker on agentic AI workflows and MCPs. In a previous life (before GPTs), he published 6 research papers in the domains of Machine Learning, Artificial Intelligence, and Natural Language Processing. He has a keen interest in B2B SaaS and ran a podcast interviewing startup founders in Japan. He can speak for hours on system design, metacognition, philosophy, and memes.

​Talk 3 - Architecting Real-World Agentic Systems

​Speakers: Jeff Yang (Solutions Architect, Noeon Research)

​Abstract: Building agentic systems for real-world domains requires more than just prompt engineering; it demands strategic decisions across both system design and model layers. This talk explores the design space of domain-specific agents, comparing rigid scaffolding, workflow patterns, and flexible agent harnesses, along with their trade-offs in maintainability and scalability. When is RAG with in-context learning sufficient, and when is fine-tuning actually necessary? By examining these architectural and model-level choices, this session provides a practical framework for making the right decisions to land agentic systems in production.

​Bio: Jeff Yang is the Lead Solution Architect at Noeon Research, where he leads the development of agentic systems designed to bridge the gap between natural language processing and formal computational logic. With a strong background in real-world AI deployment, he previously served as the Formal AI Director at CinnamonAI, spearheading visually rich document understanding pipelines and multi-use RAG systems for enterprise knowledge retrieval and QA. His expertise spans OCR, structural data extraction, and architectural patterns for building robust, domain-specific AI solutions.

​​Organizers

​​​​​​​Ilya Kulyatin is an entrepreneur with work and academic experience in the US, Netherlands, Singapore, UK, and Japan. He holds a BA in Economics, an MA in Finance, and an MSc in Machine Learning. He's a 3x founder, now helping Japan grow the local AI ecosystem through a not-for-profit community, Tokyo AI (TAI), while building an AI-native system integrator and solutions provider, Foundry Labs株式会社.

​​Supporters

​​​​Tokyo AI (​​​TAI) is the biggest AI community in Japan, with 4,000+ members mainly based in Tokyo (engineers, researchers, investors, product managers, and corporate innovation managers).

​​​Privacy Policy

​​​We will process your email address for the purposes of event-related communications and ongoing newsletter communications. You may unsubscribe from the newsletter at any time. Further details on how we process personal data are available in our Privacy Policy.