AI Shorts Factory

One sentence in, a publish-ready 9:16 video out. A single command runs seven stages — idea, script, voice, subtitles, visuals, music, render — each writing its artifact to the job folder so any stage can be re-run alone.

The subtitles are not guessed from the script: the generated voice track is transcribed back to get word-level timings, so captions never drift from the delivered audio. FFmpeg then burns them in over per-segment generated visuals with Ken Burns motion.

No stage is bound to a vendor. Voice, transcription, visuals (8 backends), music, editing and publishing each sit behind a provider interface — the same job renders as a cheap local draft or with the paid stack, from one flag.

Multi-language generation deliberately does not run twice: scripts are written per language, then the visuals are generated once and shared across all of them — image generation being the expensive stage.

~12,800 lines of Python: Typer CLI, FastAPI web UI, job queue, Pydantic config, structured JSONL logs per run, pytest + Black/Ruff/MyPy. 81 complete runs in the output archive.