17:20 ~ 18:00 Meerkat
Designing cross-product study sessions in the AI era 2026.09.03 17:20 ~ 18:00 Panda
What Android engineers review in the age of AI code reviews: Building foundations and redesigning code review
Development Process 日本語 Interpretation available
kobakenpixiv inc.
With the spread of AI coding, Android developers can now delegate parts of implementation, testing, and review assistance to AI. On the other hand, issues remain such as PR backlog, inconsistency in reviewer perspectives, and AI misinterpreting Android-specific contexts like Jetpack Compose, Edge-to-Edge, and multi-module configurations. In this session, based on our experience introducing and operating PR reviews with Gemini Code Assist in pixiv Comic Android, a large-scale product with over 10 years of history, we will share how we redesigned code reviews for the AI era. The first half covers building the foundation for sustainable AI review operations. We will introduce how project architecture guidelines written in [AGENTS.md](http://agents.md/) / Skills, Compose state management, DI, Kotlin coding conventions, and unit/E2E test setup affected the accuracy and false positives of AI reviews. In particular, we will cover Android-specific pitfalls, such as cases where AI falsely identified build errors because it could not reference generated classes for ViewBinding / DataBinding. The second half explains a review design that shifts human reviewers' responsibilities from "how code is written" to "specifications, architecture, intent, and risk assessment." We will organize practices such as avoiding reviewing minor UI tweaks and high-risk changes involving billing, authentication, or personal information with the same intensity, self AI reviews by PR authors, and perspectives that humans should ultimately judge. We will also share cases where missing specification info given to the AI led to overlooking discrepancies from specs, and cases where the AI review workflow itself became an operational cost. Rather than treating AI as an all-powerful reviewer, we will clarify the areas that can be delegated and those where humans must take responsibility, including real-world operational failures and trade-offs. For Android teams that have introduced AI coding but find code review to be a bottleneck, this session aims to provide a practical AI review foundation checklist and a draft for risk-based review redesign that you can use starting tomorrow. (Translated by the DroidKaigi Committee)
Intended audience
- Tech leads / senior Android engineers in Android teams that have adopted AI coding and are struggling with PR review backlogs or inconsistent review perspectives - Developers struggling with how to pass Android-specific context, such as Jetpack Compose and multi-module architectures, to AI reviews - Android engineers who want to design review operations defining how much to delegate to AI and where humans must retain responsibility