I participated as a developer in DevRev's panel discussion "Key Points of Future AI Technology and the Appeal of DevRev"
This page has been translated by machine translation. View original
Introduction
On May 28, 2026, I participated as a developer in a panel discussion at "AI Innovators Circle, 6th Session," an event hosted by Synthesy Co., Ltd.. This article reports on its contents.
About the Event
The event was held under the theme "How to Break Through the PoC Deadlock of Generative AI," organized by Synthesy Co., Ltd., with cooperation from DevRev, Classmethod, and the Okinawa Open Laboratory General Incorporated Association. The venue was the Synthesy office in Nihonbashi, with approximately 30 attendees. The event consisted of three segments — a keynote, a conversation, and a panel discussion — with a style that emphasized two-way dialogue.
- Official page: https://synthesy.co.jp/ai-innovators-circle06/
Structure of the Panel Discussion
The title of the panel discussion I participated in was "Key Points of Future AI Technology and the Appeal of DevRev." The four panelists were as follows.
- Moderator: Norio Shirafuji (Synthesy)
- Satohito Yamazaki (Okinawa Open Laboratory)
- Daisuke Nakashima (DevRev)
- Takumi Koshii (Classmethod / author)
I was responsible for the developer perspective, and my role was to share the hands-on feel from the standpoint of someone implementing generative AI.
What I Talked About
The panel covered three main discussion points.
The first was the difficulty of RAG. I touched on how it is not easy to retrieve the intended documents because information is significantly pruned based on semantics, and how update operations also require considerable effort.
The second was the difficulty of supervising multiple AI agents. I also shared my honest opinion that having a managed system that handles context sharing, job management, and various connections is extremely helpful, and that implementing this in-house is honestly a domain I would prefer to avoid.
The third was the difficulty of access control. I shared that achieving fine-grained access control through RAG alone is difficult, and that operational overhead tends to grow with the role design and permission policies of each service.
Discussions That Left an Impression
The discussion surrounding "data quality" was particularly memorable. Mr. Yamazaki pointed out that "at the risk of being misunderstood, once 'garbage' gets into RAG, it is difficult to remove," and Mr. Nakashima noted that "data that was once useful becomes unnecessary as time goes by." This reminded me that we need to pay attention not only to initial quality but also to freshness management. Because it was a two-way dialogue format, the discussion naturally deepened as we picked up on reactions from the audience in real time.
Closing
It was valuable to have a venue where I could share the real hands-on experience of generative AI from a developer's perspective, and I gained a great deal from it. I hope to continue speaking from a technical standpoint going forward. For readers who have opportunities to present at study groups or conferences, I encourage you to take that step forward.