Seven stages between your video and a clip worth posting.
None of this is magic and none of it is one model doing everything. Each stage does one job, and the order matters more than any single step.
-
01
Listen before you look
We fetch the audio first — never the video. Audio is a fraction of the size, and everything that decides which moments matter can be worked out from sound alone. The heavy download only happens once we know which seconds we actually want.
-
02
Transcribe every word, with timings
A transcript with word-level timestamps is the spine of the whole pipeline. Without it you cannot cut on a sentence boundary and you cannot sync a caption. Everything downstream hangs off these numbers.
-
03
Find where it is safe to cut
Shot changes from scene detection, and real pauses from the gaps between spoken words. A cut that lands mid-syllable is the amateur tell, and this is the stage that prevents it.
-
04
Choose the moments
A model reads the transcript in overlapping windows — never all at once, because a single pass reliably picks everything from the first few minutes — scores each window, then writes the strongest ones up as clips with a hook and a reason.
-
05
Repair the timings
Language models are bad at millisecond arithmetic, so we treat every timestamp they return as a rough guess and snap it to a real word boundary, then breathe outwards into the surrounding silence. This is what makes cuts land in the pause instead of the word.
-
06
Follow the subject
Faces are detected across the clip, the speaker is chosen and held, and the crop is driven by a camera that stays still inside a dead zone, ignores detections that jump implausibly far, and cuts rather than pans when the shot changes.
-
07
Finish it properly
Crop, scale, captions, loudness and encode happen in a single pass — the video is encoded exactly once. Audio is normalised to the level the platforms normalise to, and the source metadata is stripped so nothing rides along.
What we don't pretend
The virality score is an estimate from a model reading a transcript. It is useful for ranking your own clips against each other. It is not a prediction, and anyone selling you one is guessing too.
Footage with no clear subject — gameplay, screencasts, wide crowd shots — gets a blurred-background layout instead of a tracked crop, because a crop that wanders around looking for a face is worse than one that doesn't move.