Understand this story
Tetsuya Shinchi, Cloud Development Sec. Chief at the Japanese cloud infrastructure company grasys, introduced the company's AI-driven development practices. According to a report by gihyo.jp on October 26, 2026, the team routes LLMs according to task complexity: simple coding uses older-generation lightweight models to control token costs, while requirements definition and architectural challenges use models with stronger reasoning capabilities, and ChatGPT and Gemini are used in parallel. In the development process, the team adopts Issue Driven combined with AI test-driven development, inserting manual approval at the test case stage to move Human-in-the-loop earlier and prevent AI from deviating from requirements. Tetsuya Shinchi also pointed out that AI generation speed exceeds human comprehension speed, leading to problems at the review stage such as insufficient understanding and confused function responsibilities. Currently, these practices have only been disclosed by this one report, and there are no subsequent developments.
AI-generated from public reports
Latest development · Report
How grasys does AI-driven development: LLM routing, TDD test guardrails, and Human-in-the-loopDevelopments
1 confirmed developments
- Source:gihyo.jp
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 26
gihyo.jp
How grasys does AI-driven development: LLM routing, TDD test guardrails, and Human-in-the-loop
Tetsuya Shinchi, Cloud Development Sec. Chief at the Japanese cloud infrastructure company grasys, introduced its AI-driven development practices: routing LLMs by task complexity, using older-generation lightweight models for simple coding to control token costs, and using models with stronger reasoning capabilities for requirements definition and architecture challenges, while using ChatGPT and Gemini in parallel. The team adopts Issue Driven combined with AI test-driven development, inserting human approval at the test case stage to move Human-in-the-loop earlier and prevent AI from deviating from requirements. He also pointed out that AI generation speed exceeds human comprehension speed, leading to problems at the review stage such as insufficient understanding and confused function responsibilities.