Scaling Localization with AI at Lyft

What this post added

This post details the re-architecture of Lyft's localization pipeline to incorporate AI. It introduces a batch translation pipeline that uses LLMs for both translation generation ('Drafter') and quality evaluation ('Evaluator'). The 'Drafter' generates multiple candidates using prompts that include glossary terms and placeholders. The 'Evaluator' uses a reasoning-focused model to score candidates on accuracy, fluency, brand alignment, and technical correctness, selecting the best or providing feedback for iteration. This iterative critique-and-refine pattern is presented as a key innovation for improving translation quality and speed, enabling faster market launches and compliance with regulations like Bill 96.