Under the hood

The RiftCoach analysis pipeline

RiftCoach blends traditional data science with curated champion knowledge and large language models. Here is how we move from your Riot ID to an actionable coaching plan.

1. Submit and fetch

When you submit your Riot ID, RiftCoach queues a job that pulls fresh match history straight from the Riot Games API.

  • Low rate limits mean simultaneous requests can take a moment to resolve.
  • We retry intelligently so your data eventually lands even during busy hours.

2. Aggregate performance metrics

Every imported match is normalised and rolled up into a profile that highlights core combat and economy stats.

  • We track metrics like KDA, damage per minute (DPM), and creep score per minute (CS@min).
  • Outliers get dampened so a single pop-off game does not skew the entire picture.

3. Build your champion pool

We identify the five champions you played the most in your recent history to describe your live pool.

  • Right now we scan the latest 50 games to stay within rate limits.
  • Once Riot raises caps, we plan to expand to a full year of data for a richer sample.

4. Enrich with the knowledge base

Each champion is annotated with curated metadata drawn from our internal library.

  • Attributes include preferred role, class, subclass, and color theme as defined in our methodology.
  • Champions also carry traits that describe the goals they pursue and the playstyle they reward.

5. Weigh dominant themes

We infer which strategic themes define your identity by counting how often each color appears in the matches you played.

  • Main themes use a weight of 1.0 while off-themes contribute at 0.4.
  • The blend captures fringe picks without letting them overpower your core.

6. Embed and compare champions

Champion traits become one-hot vectors that we project into a fixed-length embedding space.

  • We currently rely on weighted cosine similarity to surface close fits quickly.
  • Jaccard, Dice or Tversky variants are being evaluated for future releases.

7. Match against the pro index

Your personalised embedding stand toe-to-toe with a library of high-elo player vectors.

  • We look for pros who share themes, strengths, and champion preferences.
  • The closest match becomes your pro twin and a source of next-step inspiration.

8. Generate the final analysis

Once the numbers are ready, we hand the summary off to our Claude 3 powered analyst for a human-readable breakdown.

  • Claude combines structured stats with champion lore from our knowledge base.
  • The chat feature lets you dig deeper, ask for extra tips, or explore new directions. We currently limit usage to one message per player to keep expenses in check.

What happens next?

Once analysis is complete you will be redirected to the results page. From there you can review your pro twin, explore adjacent champions, and test the chat assistant. We are continuously polishing the flow based on your feedback.