Practitioner (own blog/paper) Large-scale refactor
Google Ads uses LLMs to migrate 32-bit IDs to 64-bit in a 500M+ line codebase
An expert engineer finds locations with Code Search and Kythe; a fine-tuned Gemini toolkit generates changes that must pass unit tests; humans review and shard the changes to code owners. The manual effort had been expected to need hundreds of software-engineering years.
The numbers (as reported)
| Codebase size | 500M+ lines |
|---|---|
| Code changes fully AI-authored (landed CLs) | 80% |
| Total migration time | ~50% reduction (engineer-reported estimate) |
Facts
- Company / subject
- Tools
- Gemini (fine-tuned, internal toolkit)
- Codebase
- brownfield
- Work type
- Large-scale refactor
- Stack
- C++, Java
- Published
- 2025-01
- Last checked
- 2026-10-17
Caveats
Google uses its own model; authors call it an experience report, not a controlled study.