Sightline - one film, a million tellings

2.2 billion people live with vision impairment (WHO), and most film and video has no audio description at all. Professional AD is human-written and human-voiced - too slow and too expensive for the long tail - and where it exists, it is one fixed script: same tone, same language, same detail for everyone. Sightline rebuilds audio description as something generated, grounded, and personal. Load a film clip and its subtitle file. An open-weights vision-language model (Gemma 4 31B, served on AMD Developer Cloud's Token Factory) watches sampled frames and writes scene beats - every beat carries a confidence score and the on-frame evidence behind it, so a human reviewer can check anything the model is unsure about before it ships. A deterministic gap-fitter then places narration into the actual silences between dialogue: the model writes the words, the code owns the clock, and no narration ever talks over a line of dialogue. Then the part no fixed AD track can do: the same scene, retold. A cinephile gets visual grammar and mood. A 7-year-old gets wonder and simple words. A busy sighted listener gets a hands-free radio cut. In English, Spanish, French, German, Portuguese, Japanese, or Arabic - at a click. One film, a million tellings. Everything runs on AMD Developer Cloud through an OpenAI-compatible API, and because it is open models all the way down, the same app also runs fully offline on a laptop with a local model, for private footage or no-connection settings. Open source under the MIT license. Next: entire film libraries and live streams.