A short-video app has just shown you a clip and you watched it for 45 seconds. Will you like it, share it, comment, follow the creator, or watch it again? This system guesses — as a probability, like 0.31 — so the app can decide what to show next.
The system is four small services. Each answers one question. Click any of them — the first two run the actual model in your browser.
Type a user, a video and how long they watched. Get the probability, how confident the model is, and whether it crosses the line the business chose.
Try it →Paste a table of views and score them all. One bad row costs you that row, never the whole table — the failed rows carry a reason instead of a number.
Try it →Three numbers an operator checks after every deployment, and a chart that shows which inputs the model actually listens to (spoiler: mostly one).
Look →Is anyone calling it, is it slowing down, is it erroring, and has its output drifted? Four charts from a log the service writes about itself.
Look →Once you have a model like this, the app changes how it picks videos — and every number above was measured under the old way of picking. Does the model still work? Which of its numbers can you still trust? That is the research half of this site: RankShift, measured on real Kuaishou logs that happen to contain a randomized experiment. The one-line answer: the ranking still works; the probabilities do not, by a factor of two — and no single fix repairs both.