Built an open-source, fully local subtitle localization agent

I’ve been building SubAgent, an open-source localization system that assists with translating and editing subtitles while keeping the entire workflow on infrastructure controlled by the user. SubAgent imports SRT or RTF files, breaks the content into subtitle cues, generates multiple translation suggestions, supports human review and editing, synchronizes cues with video playback, and exports the completed subtitles. It currently supports 5 languages. The translation model runs locally through vLLM, with GoVarnam handling Romanized-to-native-script transliteration. The system rejects public inference endpoints, so subtitle text cannot be accidentally routed to a hosted model. This is the first open-source release from Hyper Latent, an AI research and product company focused on reliable, privacy-first systems. I’d appreciate feedback on two questions: Which parts of this workflow would benefit most from greater agent autonomy? Would you be willing to test the entire pipeline on my infrastructure?