CryptoChamps is a live, real-money fantasy contest platform built around crypto market data. Users draft teams of crypto assets and compete based on cumulative percentage gains or losses during short contest windows, while leaderboards update every two seconds through WebSockets. The platform also supports on-chain entries, payouts, deposits, and withdrawals.
Problem: CryptoChamps needed an AI layer for more engaging live contests
CryptoChamps already supported live crypto contests, WebSocket-based leaderboards, multiple contest formats, and on-chain transactions. However, the platform needed a richer engagement layer that could explain market movements, react to changes during a contest, and create a more dynamic experience for players.
It also needed to support users with limited market experience. The challenge was therefore to build an AI layer that could add real-time commentary and asset recommendations without slowing down gameplay or replacing CryptoChamps’ existing application architecture.
Solution: GoML built the CryptoChamps AI layer
GoML developed the CryptoChamps AI layer using Agentic AI blueprint as the best-fit delivery foundation. The system combines deterministic logic for time-sensitive operations with AI-driven workflows for live commentary, asset recommendations, and game-event processing.
The AI layer includes a game Orchestrator, Commentary Engine, Asset Team Builder, Settings Control-Plane, audio regeneration workflow, feedback and freemium controls, and supporting AWS infrastructure.
AI Matic delivery
GoML used the AI Matic Agentic AI blueprint as the delivery foundation for the CryptoChamps AI layer. The blueprint supported game orchestration, live commentary, asset recommendations, event-driven workflows, and configurable AI services while keeping the existing contest and blockchain architecture intact.
The AWS AI infrastructure includes Amazon ECS and AWS Fargate for the Commentary Engine, AWS Lambda for event-driven processing, AWS Step Functions for audio-generation workflows, Amazon DynamoDB for game-session records, Amazon S3 and CloudFront for audio storage and delivery, Amazon ECR for containers, AWS Secrets Manager for credentials, and Amazon CloudWatch for monitoring.
AI Matic delivery metrics
- Average TTFV: 20 days
- Average person-days saved: 8 days
- Avg customization: 100%
AI layer for real-time Bull and Bear commentary
The CryptoChamps AI layer turns live market activity into a broadcast-style experience through Benny the Bull and Bruno the Bear.
The Commentary Engine:
- Tracks live price movements
- Detects leaderboard and parlay events
- Identifies surges, crashes, lead changes, comebacks, and contest-end events
- Selects matching dialogue
- Adds Bull or Bear expression cues
- Attaches synchronized audio
- Pushes commentary to connected viewers through WebSockets
The original architecture also planned Amazon Bedrock with Claude for novel commentary, while routine events rely on deterministic dialogue generation.
AI layer for asset recommendations
A key part of the CryptoChamps AI layer is the Asset Team Builder, which helps users select assets for live contests.
For race contests, the AI layer:
- Fetches one-minute market data
- Scores assets using momentum, volatility, and volume
- Applies player-specific affinity weighting
- Returns the top five assets
- Provides a plain-language rationale for the recommendations
For parlay contests, the Asset Team Builder evaluates the player’s selected assets and returns an expected up, down, or flat direction with a confidence score.
AI layer for game orchestration
The AI layer also manages the lifecycle of each commentary session.
When a contest begins, the Orchestrator launches a dedicated Fargate Commentary Engine task and stores the game ID, task ARN, WebSocket URL, and status in DynamoDB. When the contest ends, the task is stopped and the game record is removed.
Each DynamoDB game-registry record carries a 24-hour TTL so stale records automatically expire.
AI layer for no-code commentary management
GoML built a Settings Control-Plane into the CryptoChamps AI layer so the CryptoChamps team can update AI-driven content and game configuration without changing application code.
The control plane supports:
- Bull and Bear dialogue
- Avatar expressions
- Voice settings
- Crypto and stock asset pools
- Contest configuration
- Audio regeneration
Configuration changes are stored in MongoDB and picked up by the Commentary Engine when the next contest starts.
AI layer for automated voice generation
The CryptoChamps AI layer includes an automated audio-generation workflow for Benny and Bruno.
- AWS Step Functions coordinates Planner, Synth, and Finalizer Lambda functions.
- ElevenLabs generates the voice clips, which are processed and stored in Amazon S3 before being delivered through CloudFront.
- This lets CryptoChamps regenerate commentary audio when scripts or voice configurations change without rebuilding the full AI layer.
AI layer for freemium controls and player feedback
The AI layer also supports usage controls for the Asset Team Builder.
- Successful recommendation calls increment usage
- Free and paid tiers are tracked per user
- Free users are restricted after reaching their configured limit
- The default free-use limit is three recommendations
- Rejected requests do not consume another use
Player thumbs-up and thumbs-down feedback is stored with game records, giving CryptoChamps a signal for improving ranking rules and commentary over time.
AWS infrastructure supporting the CryptoChamps AI layer
GoML built the CryptoChamps AI layer using a distributed AWS architecture that separates commentary, recommendations, settings, audio generation, and orchestration workloads.
The technology stack includes:
- Amazon ECS and AWS Fargate
- AWS Lambda
- AWS Step Functions
- Amazon API Gateway
- Application Load Balancer
- Amazon DynamoDB
- Amazon S3
- Amazon CloudFront
- Amazon ECR
- AWS Secrets Manager
- Amazon CloudWatch
- MongoDB Atlas
- Terraform
- GitHub Actions
Security and resilience of the CryptoChamps AI layer
The AI layer uses private AWS networking and service-level access controls.
- Fargate tasks and Lambdas run inside private VPC subnets
- Secrets are stored in AWS Secrets Manager
- Dedicated IAM roles follow least-privilege access
- NAT controls access to external APIs
- CloudWatch captures service logs
- Price-feed fallbacks keep commentary running when a primary provider fails
The technical documentation notes that the admin Settings endpoints still need authentication restrictions before production.
Quality assurance for the CryptoChamps AI layer
Testing for the AI layer covers:
- Game provisioning
- Commentary generation
- Race and parlay events
- Market-data fallbacks
- Asset recommendations
- Freemium limits
- Settings validation
- Audio regeneration
- External-service failures
- Game cleanup
The project success criteria include commentary delivery within two seconds of a game event, plain-English reasoning for Asset Team Builder selections, correct freemium enforcement, and complete feedback capture for AI-assisted team-building sessions.
Impact
- <2 seconds target for live Bull and Bear commentary
- 100% target feedback capture across AI Team Builder sessions
- Top 5 ranked asset recommendations in race modes
- 3 default free Asset Team Builder uses
- 24-hour automatic TTL for stale game-session records
- $50–$100/month target Bedrock budget at MVP traffic levels
About
Before Gen AI and after Gen AI
“With CryptoChamps’ AI layer, we brought real-time commentary, intelligent asset recommendations, and automated game workflows together without disrupting the platform’s existing architecture.”
Prashanna Rao, Head of Engineering, GoML.
Key takeaways from the CryptoChamps AI layer
Common challenges when building an AI layer
- AI processing must not interrupt live gameplay
- Market recommendations need explainable logic
- Commentary needs low-latency delivery
- External APIs require fallback mechanisms
- AI content needs to be configurable without constant deployments
Practical guidance for building an AI layer
- Keep high-frequency decisions deterministic
- Use generative AI selectively for richer language
- Separate AI services from core transactional systems
- Pre-generate audio instead of waiting for TTS during gameplay
- Add price-feed fallbacks for resilience
- Capture player feedback to refine future AI behavior
Build your AI layer with GoML
Partner with GoML to build an AI layer that adds real-time intelligence, recommendations, automation, and generative experiences to existing applications, powered by the AI Matic platform.
