SberDevices has implemented a frontend agent system that combines deterministic workflows with AI-driven agent loops. The system uses the GigaChat-2-Max model to perform automated code reviews and analyze CI/CD job failures within GitLab.
The AI code reviewer operates on a server-side TypeScript and NestJS architecture. It automatically processes merge requests by analyzing git diffs, filtering out service files, and providing specific feedback on logic errors, null/undefined handling, and performance issues.
Between January and July, the AI reviewer processed over 4,000 merge requests across 90 projects, identifying potential issues in 8% of cases. Over 90% of the bot's comments highlight specific problems rather than general recommendations.
The company also deployed a notification bot named Chuck Norris to reduce delays in the review process. It sent nearly 4,000 reminders across 51 projects between January and August, with most merge requests requiring only one ping.
For CI/CD troubleshooting, a custom parser extracts relevant log segments to provide context to GigaChat, which then identifies the cause of failure and suggests fixes. The system also supports automated repairs for common build, Docker, and Nginx configuration errors.
