saitejamummadi/subagent

SubAgent is a subtitle localization tool meant to run entirely inside a private network, with no data ever sent to an outside company. You import a subtitle file in SRT or RTF format, optionally attach the matching video, and then go through each subtitle line, called a cue, translating it into Telugu, Hindi, Tamil, Malayalam, or Kannada. As you work, a language model running on your own graphics card suggests translations you can accept or edit, and the finished subtitles can then be exported. The project exists because most subtitle translation tools send your dialogue to a third party API, which is a problem if the content is under a non disclosure agreement or if you simply do not want an outside vendor involved. SubAgent keeps the editor, the translation model, and the conversion between romanized and native script text all inside infrastructure you control, with no calls to any hosted AI service anywhere in the code. The editor shows the video on one side and the list of cues on the other, with each cue expandable to show the original text, an editable translation, and suggestion chips with confidence scores from the model. Behind the scenes, a React frontend talks to a FastAPI backend backed by SQLite, which in turn talks to a translation model served through vLLM and a transliteration service, both reachable only over private or local network addresses so that pointing the system at a public server on purpose still fails. Running it requires Windows, Python 3.12, Node 20 or newer, Docker Desktop with GPU support, and an NVIDIA graphics card for live translation suggestions, though the editing and export features work without a GPU using a stand in translation worker. The project includes a one command startup script as well as instructions for starting each piece by hand. It is licensed under MIT.