Rift AI Scoreboard

AI enhanced Scoreboard for Rift that includes Guru feedback features based on automated labeling and ML training for increased awareness for player improvement.

Comment

Inspiration

What it does - Features

Clutch Moment Detection

  • ML-powered detection using 9 engineered features
  • 30-second sliding window analysis
  • Validation AUC ≥ 0.80

Narrative Generation

  • AWS Bedrock integration with Claude models
  • Personalized, engaging narratives
  • A/B testing framework for model optimization

A/B Testing

  • Compare different Claude models (Haiku vs Sonnet)
  • Track cost, latency, and quality metrics
  • QuickSight dashboards for experiment analysis

How we built it

Challenges we ran into

QuickSight doesn't support DynamoDB as a direct data source through the API, so I needed to use Athena to query DynamoDB instead. I created an Athena data source that queries DynamoDB through a federated query connector. I set up S3 as an intermediate data source, to handle the export process from DynamoDB to Athena.

Accomplishments that we're proud of

Getting the Riot API key to work and pull match data that was then used to train a model with automated labeling. Creating

What we learned

1. Data Collection & Processing

  • ✅ Integrated with Riot API (Account-V1 + Match-V5)
  • ✅ Collected 20 real ranked matches
  • ✅ Extracted 362 clutch candidate windows
  • ✅ Automated labeling using heuristic rules (249 clutch, 113 non-clutch)

2. Machine Learning Model

  • ✅ Trained LightGBM binary classifier
  • ✅ Validation AUC: 1.0000 (perfect on dataset)
  • ✅ Model uploaded to S3: s3://riftguru-models-643236078550/clutch_lgbm.pkl
  • ✅ 9 features: kills, objectives, CC time, positioning, team tactics

3. AWS Infrastructure

  • ✅ S3 buckets for matches and models
  • ✅ DynamoDB table with 10 sample clutch moments
  • ✅ 9 Lambda functions deployed
  • ✅ CloudWatch logging enabled

4. Demo Data

  • ✅ 10 clutch moments loaded in DynamoDB
  • ✅ AI-generated narratives for each moment
  • ✅ Impact scores ranging from 0.85 to 0.94
  • ✅ Ready for visualization

What's next for Rift AI Scoreboard

Pull larger amounts of match data and retrain the model. Enhance features to offer more detailed analysis and coaching for players.

Share this project:

Updates