Technical Architecture
Last updated
Last updated
Billy Bets has been built utilizing a combination of the and proprietary coding. This allows Billy Bets to operates as a fully autonomous agent which leverages advanced AI models, bespoke sports datasets, and intelligent social interaction capabilities. This technical overview details the core components and their interactions.
The Billy Bets character has been hand crafted to replicate the personality of top sports betting personalities, both real and fictional. The agent is able to pull from a number of pop-culture and sports references to provide a geniunely entertaining personality with sacrastic wit, allthewhile providing deep sports analysis.
Billy Bets leverages multiple state-of-the-art AI models including Anthropic’s Claude 3.5, OpenAI’s ChatGPT 4o, Meta's LLaMA 3.2, and Perplexity for different specialized functions. Each model serves specific purposes within the system, from content generation to complex decision-making processes.
This has been fine-tuned using a combination of Reinforcement Learning from Human Feedback (RLHF) and supervised training techniques.
Supervised Fine-Tuning: Manually curated conversations and replies, emphasizing Billy’s tone and character, is used to train the model. This step is important to ensure that our agent maintains its funny and degenerate style across various interaction scenarios.
Reinforcement Learning: A reward-based feedback loop is implemented, where responses are scored based on engagement metrics from X (likes, retweets, comments) to optimize for more engaging content.
Billy Bets implements a multi-layered data processing system pulling from relevant sources that ensure it to respond in real time with context, relevance and historical knowledge allowing for it to create organic engagement with users.
Key integrations include:
Sports Data Integration through Sportsdata.io API with real-time websocket connections
Blockchain Oracle Integration via Polymarket's on-chain APIs
Persistent Storage Layer utilizing Supabase for distributed memory management to support RAG processing.
Deployment Infrastructure through Render for scalable service management
Daily Betting Intelligence via the SportsTensor API
A sophisticated RAG-based memory architecture has been implmented that allows for:
Vector-based conversation storage for efficient context retrieval
Hierarchical user profile management system
Distributed sports knowledge base with real-time updates
Context-aware memory pruning and optimization
