What existed
Short-form video discovery runs on engagement signals: watch time, taps, shares, the loudness of a crowd. Volume decides what you see next. How a video actually feels is nowhere in the loop, so a feed cannot tell the difference between loud and good.
What it became
VISION classifies short-form video by its emotional quality, using frontier video-understanding models that watch the way a person does. Recommendation follows that reading, so discovery follows vibe rather than volume. It runs in production under VALA.
What was built
- A video-understanding pipeline on frontier models
- Emotional classification of short-form video, beyond engagement signals
- A recommendation engine driven by that classification
- Production systems, under VALA
What it shows
The thing is not only made, but smart enough to find its people. That is the intelligence underneath, and it is built here.