How the AI coaching works

What is AI-generated, and what is not

Every number you see — CS per minute, gold difference at 10 minutes, win rate against a champion class — is computed by ordinary code from your match data. Nothing about those figures is generated, guessed, or estimated by a model.

The words of a coaching review are AI-generated. A large language model is given the numbers that have already been computed, along with labels for what they mean, and writes a short review about them. It never does the arithmetic itself and it never sees your raw match data.

That split is deliberate: it means you can check the numbers against the match they came from, and judge the wording separately.

Which model, and where it runs

Reviews use Groq and the openai/gpt-oss-120b model. Our application and database are hosted on AWS in Paris. Groq is a separate provider: AI processing is not restricted by our AWS region, and we do not promise Paris-only or EU-only inference.

Groq may retain request data for reliability and abuse monitoring under its data controls. Its published policy places retained customer data in the United States. See Groq's data policy for the provider's retention controls. We also use Riot and OP.GG for game information; they are separate services from our Paris infrastructure.

Three things it gets wrong

Worth knowing before you take any of it to heart:

1. It can be confidently wrong. Language models write fluently whether or not they are right. Assured, well-organised prose is not evidence of a correct diagnosis.
2. It only sees 15 games. That is a small sample. A “pattern” over 15 games can easily be variance — a rough week, a run of hard matchups, or nothing at all.
3. It does not watch replays. It works from per-minute timeline data and end-of-game statistics, not from what happened on screen. It cannot see a mis-click, a lag spike, a bad call from a teammate, or the reason you made a decision that looks wrong in the numbers.

It is advice, not a verdict

A review is a suggestion for practice. You read it and decide what to do with it. Nothing happens automatically as a result: no score is assigned to you, no ranking is affected, no profile of you is built and carried anywhere else. If a review is wrong, disagree with it — the statistics it is built from are on the same page and are checkable against the match.

More detail

The full system description, including the risk classification and what we consider still outstanding, is kept in the repository at docs/compliance/ai-system-description.md. What we hold about you, and how to have it deleted, is on the privacy page.