Scaling Code Translation

Legacy systems written in obsolete programming languages are costly and risky to modernize using manual translation methods. While large language models (LLMs) show promise for their use in code translation, their effectiveness decreases as system scale increases. This poster presents an incremental Ada-to-C++ translation workflow that augments LLMs with static analysis, context-aware prompt generation, and automated glue code. The approach preserves the translation state across increments and keeps developers in the loop for quality control. Early results show substantial reductions in translation errors and success rates of 87% to 100% in key error categories, with a target of achieving a fourfold improvement in translation speed.