🎮 RiftGuru - AI-Powered Clutch Moment Detection

Automatically detect and highlight the most impactful moments from League of Legends matches using machine learning

🌟 Overview

RiftGuru uses hybrid rule-based and ML-augmented detection to identify clutch moments in League of Legends matches. Built for the AWS AI Hackathon 2025, it processes Riot API data through a serverless pipeline to generate AI-powered narratives of your best plays.

View Live Dashboard | Watch Demo Video | Read Documentation

✨ Features

• 🤖 ML-Powered Detection: LightGBM classifier with 1.0 validation AUC
• ⚡ Serverless Architecture: AWS Lambda + DynamoDB + S3
• 🎯 Real Riot API Data: Processes actual ranked match timelines
• 📊 Beautiful Dashboard: Vercel-hosted visualization
• 🔄 Automated Pipeline: End-to-end data collection to visualization
• 📝 AI Narratives: Bedrock-generated descriptions of clutch plays

🏗️ Architecture

Riot API → Lambda (Extract) → Lambda (Predict) → Lambda (Enrich) → DynamoDB → Dashboard
              ↓                    ↓                    ↓
            S3 Matches         ML Model (S3)      Bedrock AI

Key Components

Component	Technology	Purpose
Data Collection	Riot API + Python	Fetch match timelines
Feature Extraction	AWS Lambda	Extract 30s windows with 9 features
ML Model	LightGBM	Binary classification (clutch/non-clutch)
Narrative Generation	Amazon Bedrock	AI-powered play descriptions
Storage	DynamoDB + S3	Clutch moments + raw data
Frontend	HTML/JS + Vercel	Dashboard visualization

🚀 Quick Start

Prerequisites

• AWS Account with CLI configured
• Python 3.12+
• Node.js 18+ (for CDK)
• Riot API Key

Installation

# Clone repository
git clone https://github.com/YOUR_USERNAME/riftguru.git
cd riftguru

# Install Python dependencies
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# Install CDK
npm install -g aws-cdk

# Set up environment
export RIOT_API_KEY="your-api-key"
export AWS_REGION="us-east-1"

Deploy Infrastructure

# Bootstrap CDK (first time only)
cdk bootstrap

# Deploy stack
cdk deploy RiftGuruClutchPlayStack

Collect Data & Train Model

# Collect matches
python scripts/collect_match_data.py

# Extract windows
python scripts/extract_all_windows.py

# Auto-label data
python scripts/auto_label_windows.py

# Train model
python scripts/train_model.py

# Upload model to S3
python scripts/upload_model_to_s3.py \
  --model models/clutch_lgbm.txt \
  --bucket riftguru-models-YOUR_ACCOUNT_ID

Load Demo Data

# Quick demo with sample data
python quick_demo_test.py

View Dashboard

Open the live dashboard or run locally:

# Open in browser
open dashboard/index.html

📊 Results

Metric	Value
Matches Collected	20
Windows Extracted	362
Clutch Moments Detected	249 (68.8%)
Model Validation AUC	1.0000
Training Time	< 1 second
Model Size	0.01 MB
End-to-End Latency	< 5 seconds

🧠 Machine Learning

Features (9 total)

1. kill_count: Kills in 30s window
2. death_count: Deaths in window
3. damage_taken: Damage dealt to enemies
4. damage_received: Damage received
5. objective_flag: Major objective secured (Baron/Dragon/Tower)
6. crowd_control_time: Total CC duration
7. position_score: Distance to enemy territory (0-1)
8. role_score: Encoded role (1-5)
9. team_tactic: Team coordination indicator

Model Performance

• Algorithm: LightGBM Binary Classifier
• Training Data: 362 labeled windows (289 train, 73 validation)
• Validation AUC: 1.0000 (perfect on dataset)
• Optimal Threshold: 0.6855
• F1 Score: 1.0000

Labeling Strategy

Automated heuristic-based labeling:

• Multi-kill (3+ kills in 30s)
• Objective with kills
• High damage + objective
• Extended CC + kills
• Team coordination + kills

🏛️ AWS Infrastructure

Services Used

• Lambda: 9 functions for pipeline stages
• DynamoDB: Clutch moments storage
• S3: Match data + ML models
• Step Functions: Pipeline orchestration
• CloudWatch: Logging and monitoring
• Bedrock: AI narrative generation
• CDK: Infrastructure as Code

Cost Estimate

• Development: ~$5/month (free tier eligible)
• Production: ~$20/month (1000 matches/day)

📁 Project Structure

riftguru/
├── lambda/                 # Lambda function code
│   ├── extract_windows/   # Window extraction
│   ├── predict_clutch/    # ML prediction
│   ├── enrich_narratives/ # Bedrock integration
│   └── store_clutch_moments/ # DynamoDB storage
├── scripts/               # Training & data collection
│   ├── collect_match_data.py
│   ├── auto_label_windows.py
│   ├── train_model.py
│   └── upload_model_to_s3.py
├── dashboard/             # Frontend
│   └── index.html
├── models/                # Trained models
├── data/                  # Training data
├── clutch_play_stack.py   # CDK infrastructure
└── app.py                 # CDK app entry point

🔧 Configuration

Environment Variables

# AWS
AWS_REGION=us-east-1
AWS_ACCOUNT_ID=your-account-id

# Riot API
RIOT_API_KEY=your-api-key

# Model
MODEL_S3_BUCKET=riftguru-models-{account-id}
MODEL_S3_KEY=clutch_lgbm.pkl
CLUTCH_THRESHOLD=0.75

📖 Documentation

• Architecture Flow
• Deployment Guide
• Known Issues
• API Documentation

🎯 Roadmap

• Fix damage feature extraction
• Collect 5000+ training samples
• A/B test Claude Haiku vs Sonnet
• Real-time detection during live games
• Mobile app
• Multi-region support

🤝 Contributing

Contributions welcome! Please read CONTRIBUTING.md first.

📄 License

MIT License - see LICENSE file

🙏 Acknowledgments

• Riot Games for the API
• AWS for the infrastructure
• Amazon Bedrock for AI narratives
• LightGBM team for the ML framework

📧 Contact

• Author: David Anderson
• Email: info [at] homesolarsavemoney [dot] com
• GitHub: @fapulito
• Demo: Live Dashboard

---

Built with ❤️ for the AWS AI Hackathon 2025