Veldora
Not releasedLong video in, shorts out, on your own machine. Three hours of podcast becomes eleven vertical clips worth posting, without a single frame leaving your GPU.
- AI
- Whisper · NLLB · pyannote
- Stack
- WinUI 3 · Python sidecar
- Requires
- NVIDIA GPU
- Year
- 2026
Problem
Every tool that turns long video into shorts wants you to upload the footage first, which means a render queue, an account, and handing your unreleased work to somebody else's server.
Solution
Veldora keeps a Python AI engine running beside the app with every model resident in VRAM. Whisper handles the speech, NLLB the translation, pyannote the speakers. It scores every second of the transcript for hooks, punchlines and story beats, then cuts vertical clips with burned-in subtitles. No account, no upload, no queue.
Architecture
A WinUI 3 app talking to a local Python engine on 127.0.0.1 that holds Whisper, NLLB and pyannote in VRAM, with FFmpeg and the Python runtime bundled so the user installs nothing. The app reads your GPU, VRAM, RAM, cores and free disk, then recommends the model sizes that will actually fit.
Outcome
The app and the AI engine are finished and working. The installer is the one thing left. Shipping a bundled Python runtime and a multi-gigabyte model set as a single Windows install is its own project, and it's the one I'm on.