Plataforma de Análise de Performance de Atletas
ellttdave/athlete-performance-platform-showcase
Showcase de arquitetura LLM + RAG + MCP em TypeScript
Ver no GitHubMCP server TypeScript que demonstra integração avançada de LLM com Claude, sistema RAG com embeddings vetoriais e arquitetura full-stack moderna. Implementa padrões de produção para chamadas estruturadas a ferramentas, processamento de documentos com Google Document AI e busca semântica em PostgreSQL com pgvector.
Instalação
npx convez-ai ellttdave/athlete-performance-platform-showcase1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Arquitetura de referência para integração LLM em produção
- Implementação de padrão RAG com vector search e PostgreSQL
- Processamento de documentos PDF com extração inteligente
- Padrão MCP customizado para separação de concerns
Veredito

Projeto técnico sólido que serve como blueprint para aplicações LLM em produção, com implementação genuína de RAG, processamento de documentos e padrão MCP. Recomendado para quem busca referência de arquitetura escalável, não para uso direto em produção sem customização.
README
A tradução para o português deste README ainda está na fila.
Athlete Performance Analytics Platform - Technical Showcase
A production application demonstrating advanced LLM integration, RAG systems, and modern full-stack architecture patterns.
🎯 Project Overview
This repository showcases the technical architecture and implementation patterns from a production application that integrates:
- LLM Integration (Anthropic Claude) with structured tool calling
- RAG System (Retrieval-Augmented Generation) with vector embeddings
- MCP-like Architecture for clean separation of concerns
- Modern Full-Stack development with Next.js, TypeScript, and PostgreSQL
Note: This repository contains sanitized code examples and architecture documentation for showcaase purposes. The full production application remains private to protect intellectual property.
🏗️ Architecture Highlights
MCP-Like Tool Integration Pattern
A custom architecture pattern that mimics Model Context Protocol (MCP) to provide clean separation between LLM communication logic and data functions.
Key Benefits:
- Single gateway for all LLM tool calls
- Easy migration path to real MCP server
- Extensible tool registry pattern
- Comprehensive error handling
RAG Implementation
A production-ready RAG system using vector embeddings and semantic search.
Key Features:
- Google Document AI for advanced text extraction (Form Parser, Layout Parser)
- Automatic PDF splitting for large files (>15 pages or >4MB)
- OpenAI embeddings (text-embedding-3-small)
- pgvector for PostgreSQL vector similarity search
- Vercel Blob Storage for document persistence
- Individual document re-processing to avoid timeout issues
- Fallback mechanisms (unpdf) for reliability
- Automatic knowledge base integration
- Top-K retrieval with similarity scoring
Architecture Note:
- RAG is used for general knowledge, research articles, and contextual explanations
- Precise numeric/reference data is stored in PostgreSQL (not RAG) for deterministic queries
Full-Stack Design
Modern web application architecture with:
- Frontend: Next.js 14, React 18, TypeScript, Tailwind CSS
- Backend: Next.js API Routes, Server Actions
- Database: PostgreSQL (Neon) with Prisma ORM, pgvector extension
- AI/ML: Anthropic Claude API, OpenAI Embeddings
- Document Processing: Google Document AI, pdf-lib for PDF splitting
- Storage: Vercel Blob Storage for document persistence
- Deployment: Vercel
📚 Documentation
- Architecture Overview - System design and architecture patterns
- MCP Architecture - Tool registry pattern and implementation
- RAG Implementation - Vector embeddings and semantic search
- LLM Integration - Tool calling and structured outputs
- Tech Stack - Technology choices and rationale
💻 Code Examples
Sanitized code examples demonstrating key patterns:
- MCP Router Pattern - Tool registry and routing
- RAG Implementation - Vector embeddings and search
- LLM Integration - Tool calling examples
🛠️ Key Technical Features
1. Custom Tool Registry Pattern
Clean separation between LLM communication and data functions:
// Extensible tool registry
const TOOL_REGISTRY = {
'analyze_data': {
description: 'Comprehensive data analysis tool',
handler: async (params) => { /* ... */ }
}
}
2. RAG System with Vector Search
Production-ready retrieval-augmented generation:
- Vector embedding generation
- Semantic similarity search
- Knowledge base integration
- Context retrieval for LLM
3. Structured LLM Outputs
Type-safe JSON generation with validation:
- Schema validation
- Error handling
- Data integrity measures
- Reliable structured outputs
4. Production-Ready Patterns
- Comprehensive error handling
- Logging and monitoring
- Type safety with TypeScript
- Scalable architecture
💡 Technical Insights
Why MCP-Like Architecture?
- Provides clean separation of concerns
- Makes LLM integration maintainable
- Allows easy addition of new tools
- Facilitates migration to real MCP server
RAG Implementation Decisions
- Chose pgvector over external vector DB for simplicity
- OpenAI embeddings for reliability
- Top-K retrieval for performance
- Integration with existing PostgreSQL infrastructure
LLM Integration Challenges
- Structured output reliability
- Data integrity enforcement
- Error handling and fallbacks
- Token optimization
🚀 Key Achievements
- ✅ Implemented custom MCP-like architecture for LLM tool integration
- ✅ Built production RAG system with vector embeddings
- ✅ Designed scalable data models for multi-modal analytics
- ✅ Created evidence-based analysis engine with citation requirements
- ✅ Deployed production application with optimized performance
📄 License
Technical showcase - Code examples are for demonstration and educational purposes.
Note: This repository contains sanitized code examples and architecture documentation. The full production application remains private to protect intellectual property.
Connect
- LinkedIn: http://linkedin.com/in/david-elliott-6304555a
- Email: ellttdave218@gmail.com