Gemini · ADK · ClickHouse MCP

Every platform can show you the minute an audience stops watching. None of them can tell you whether the scene dragged or the stream stuttered, and those two demand opposite fixes. Walkout reads 13 million playback events, then goes and watches the film at that exact timecode.

Read the docs
Live cluster

Retention forensics

Pick a title and follow it down: where its audience walked out, what was on screen when they did, and why, because a scene that drags and a stream that stutters make the same shape on a chart and need opposite fixes.

1

Where they left

Every drop below is a moment when far more people quit than this title normally loses. Not the gentle drift every film has: the cliffs.

Hazard rate across the runtime

Out of everyone still watching, the share who quit in each ten seconds. Shaded bands are the cliffs. The end credits are left out, because leaving there means you finished.

still watching quitting cliff
2

What was on screen

Jump the film to any cliff and watch it yourself, then ask Gemini what it sees. It is never told that anyone left there, so when the picture agrees with the numbers, that agreement means something.

Cueing the film
00:00:00

Ask Gemini to watch it

Reading 00:03:40–00:04:10

Pick a cliff on the left, then press the button. Gemini watches those exact seconds of the real film and reports what is happening: the pacing, the dialogue, anything visibly broken.

3

Why they left

The numbers say who left; the film says what they were watching. Put the two together and the fix has an owner: an editor, a translator, or an engineer.

Method

Two witnesses that never spoke to each other.

Your telemetry saw the audience but never the film. A model can watch the film but has no idea which thirty seconds matter. Walkout is the conversation between them.

01

Where they left

Thirteen million play, pause and quit events. We keep only the moments where far more people left than this film normally loses.

Three cliffs, to the second
02

Who left

The same moment, cut by phone, TV, app build, country and language. Everyone at once points at the screen. One group points at whatever they had in common.

A cohort, or nobody
03

What was on screen

Gemini watches those exact seconds. It is never told that anyone left there, so when it flags the same moment anyway, that is a second witness.

A reading of the footage
04

What to fix

The numbers and the footage either agree or they do not. Either way you get one verdict, and the name of whoever should act on it.

A verdict, and its owner

This demo runs on Sintel (Blender Foundation, CC-BY) with 13 million simulated viewing events and cliffs planted at known timecodes, so the diagnosis can be graded rather than admired. The film is streamed straight from YouTube, never copied here.

Full documentation→ How each verdict is decided, and what it refuses to guess at.

Gemini 3.5 Flash Gemini 3.6 Flash · video Agent Development Kit ClickHouse MCP server ClickHouse Cloud · 13.07M events FastAPI