Scrub your own replay
In-game, from the cameras the game gives you, hunting for a mistake you already failed to notice once. Most people watch two minutes and quit.
Rocket League · AI replay coaching
Upload a replay. AI coaching built on your own match data tells you what went wrong — and every claim is a link into the exact moment, in 3D.
The scene is rebuilt from the replay’s own telemetry at twenty ticks a second. That is precisely why the camera is yours.
What you don’t have is a camera that goes where you want, and somebody to tell you which three seconds out of five minutes actually cost you the game.
In-game, from the cameras the game gives you, hunting for a mistake you already failed to notice once. Most people watch two minutes and quit.
Priced by the hour and scheduled around somebody else’s calendar. Genuinely good — and you will do it once a month, not after every session.
Eight deterministic detectors read the telemetry, the five clearest mistakes get written up, and every sentence is a link into the exact tick range that produced it.
And when there is nothing worth saying, it says nothing. That is the part that makes the rest worth reading.
These are real claims from a real report, each bound to a real tick range. Pick one and the scene above jumps to it, camera already positioned to make the claim legible. That chain — claim, moment, camera — is the product.
Drag the .replay file straight from your Demos folder. Parsing, detection and the report take under a minute.
Never forty. The report shows the eight most severe moments, and the model may write at most five claims about them.
Swipe the flagged moments on your phone, mark each one fair or not, and take a single thing into the next session.
Every moment comes from a deterministic geometric detector. The same replay always produces the same moments — forever. The model never sees the game; it only narrates what these measured.
You and a teammate close on the same ball at the same time while the opposition contests it. One of you spent a challenge for nothing.
You are the deepest player, you are up the field, and the opposition takes the ball while nobody is goal-side of you.
You challenge as last man at a moment when holding — decelerating, staying between ball and goal — was geometrically available.
Empty in your own third while the opposition has the ball. Being out of boost is not a mistake. Being out of boost there, then, is.
A teammate is first man and you are sitting ahead of the ball, so there is no outlet and no cover if the challenge goes wrong.
You spent a flip, a jump or a chunk of boost, and two seconds later nothing had changed. The resource is gone and it bought nothing.
Both players committed, or the non-taker grabbed a mid pad instead of back boost on a kickoff that was going to be lost anyway.
Getting demoed is bad luck. Getting demoed as the only player between the ball and your net is usually position and awareness.
If fewer than two moments clear the bar you get no report at all — not a padded one. A tool that can never say “clean game” is a tool you cannot believe when it says otherwise, so refusing to write is a feature and it is tested like one.
The same discipline runs through the writing. A claim may only state numbers that appear in the evidence of the moments it cites, and that rule is enforced in code before anything is stored — not requested in a prompt. If the model invents a figure, the report is thrown away and regenerated.
Don’t take our word for it. Every claim on this site is one tap from the tick range that produced it.
Fewer than two moments cleared the bar, so there is nothing to show you. That is the honest answer, not an empty page.
The viewer is the thing nobody else has, so it is not the thing to gate. What costs money is volume and the longitudinal features a serious player actually needs.