ARVEN
Agentic video intelligence.
A video-remix engine that finds the cuts on its own. A trained model detects scene boundaries, then a cut, assemble, and caption pipeline turns raw footage into finished short-form video, all wrapped in a hardened multi-tenant shell.
The ARVEN video-intelligence console.
Cutting raw footage is the slow, expensive part of making video ads. ARVEN automates it. A two-pass detector fuses seventeen signals with trained gradient-boosting models to decide where one scene ends and the next begins, and it does that on your own hardware with no per-minute cloud bill.
It is the deepest engineering in the suite and the largest by far. The interesting claim is not that it works, but that its accuracy is measured and locked down: the model scores about 0.95 on an F1 benchmark, and that number is protected by tests that run without shipping any sample video.
- Scene detection that fuses seventeen signals with trained gradient-boosting models, measured at roughly 0.95 F1, with no per-minute cloud cost.
- A cut, assemble, and caption pipeline with silence-stripping and batch export.
- Audio intelligence: transcription, vocal isolation, and reading text off the screen.
- A hardened multi-tenant shell with dual sign-in (token and API key), per-route credit billing, and app-wide two-factor authentication.
- A 170-command line interface that mirrors the API, with a build gate that fails if any route is missing its command.
A Node and Express backend is organized as two dozen dependency-injected route modules, with a React console on top and a Python machine-learning core underneath holding the trained models. Scene-detection accuracy is pinned by golden-fixture tests that carry the decision data but not the raw footage, so the model can be verified the same way every time.
- Node
- Express
- React
- Python
- scikit-learn
- OpenCV
- Whisper
- FFmpeg
- better-sqlite3
Real figures from the build, not marketing math.