Yogiyo Reveals Case Study on Building Automated Pipelines as 'Verification' Becomes the New Bottleneck in the AI Coding Era

Yogiyo's Tech Blog emphasizes that in environments where AI writes code, the developer bottleneck has shifted from 'writing code' to 'verification'. Through case studies of the new game service 'YogiWorld', a pilot project for Product Engineers, and the 'Order Detail Big Size Ad' in iOS and Android apps, it discusses the importance of verification that humans must perform after AI's rapid code generation.

YogiWorld: Building an Automated Verification Pipeline

In YogiWorld development, an automated verification pipeline was built to maintain quality while merging approximately 240 PRs over two and a half months. The CI gate consisted of approximately 1,100 frontend unit tests, approximately 200 backend unit tests, approximately 150 integration tests, and 36 E2E tests. Notably, visual regression testing adopted a method that does not store baseline images but instead renders code before and after changes on the same runner for each PR to compare pixel differences, thereby reducing noise caused by environmental differences.

The development cycle is divided into a 'develop' stage for agreeing on direction and a 'ship' stage for proceeding from the implemented branch to deployment. Humans are responsible only for agreeing on direction and confirming final intent, while the pipeline handles exploration, implementation, review, and deployment. Through this pipeline, YogiWorld, officially launched on July 2, recorded over 48,000 cumulative plays by September 21.

Big Size Ads: Verification on Unfamiliar Platforms

In ad projects requiring native app code modifications, MCP (Model Context Protocol)-based automation was utilized to connect building, running, screen measurement, and fixing into a single loop. The 'bug where ads refresh every 10 seconds' mentioned at the beginning could not be confidently resolved with AI-suggested fixes alone, but was solved by confirming the reproduction path down to file and line number and directly verifying the symptom in the simulator. Additionally, duplicate request bugs were found by inserting logs into the ad loading pipeline to count request numbers.

Conclusion: Designing Verification Methods is the Core Competency

The faster AI writes code, the faster incorrect fixes can be repeated. Therefore, the tasks humans must handle are agreeing on direction, limiting scope, and the judgment to question premises when detours are repeated. Based on these experiences, Yogiyo is developing verification tools such as 'RECORDYO', which records and replays actual app network sessions, and is working on porting the visual regression testing engine to other repositories. The role of a Product Engineer in the AI era is defined not as writing more code, but as designing verification methods that provide confidence in deployment.