Lyft built an AI-driven localization pipeline to replace slow manual translation workflows. The system uses a dual-path batch architecture where LLMs generate draft translations in parallel with a translation management system (TMS) for human oversight. A Drafter-Evaluator separation improves quality by decoupling generation from assessment. Context injection and deterministic guardrails handle brand, legal, and regional constraints. About 95% of translations pass human review with minimal changes, reducing turnaround from days to minutes. A separate low-latency pipeline handles real-time use cases like ride chat.
Sort: