How we improved translation experience with cost efficiency

This post details the evolution of Grab's translation service by implementing on-the-fly language detection using Lingua, with a fallback to a third-party service. It describes the process of training a custom lightweight translation model using synthetic data from a larger LLM, incorporating domain-specific context and a post-translation quality check to mitigate number mutation issues. Caching strategies and optimizations for system messages were also introduced to reduce costs.