FL Platform - Decentralized AI Orchestration

A hybrid federated learning system that bridges the gap between local browser-based simulation and real-world distributed training. Orchestrated by Google Gemini, visualized with React 19, and powered by FastAPI.

🚀 Live MCP Server: Hugging Face Space

Key Features

🌐 Browser-Based Training (WebGPU)

• Zero-Install Client: Train models directly in the browser using WebGPU compute shaders.
• LoRA Support: Efficient fine-tuning of Large Language Models (LLMs) using Low-Rank Adaptation (LoRA).
• Privacy-First: Raw data never leaves the user's device; only weight deltas are shared.
• OPFS Checkpointing: Persistent storage of model weights and training progress using the Origin Private File System.
• Sybil Resistance: Device fingerprinting prevents malicious actors from flooding the network with fake clients.

🤖 Gemini AI Orchestrator

• Autonomous Script Generation: Gemini 1.5 Pro dynamically writes PyTorch/Flower training clients based on project requirements.
• Model Discovery: Automatically finds optimal models on Hugging Face based on task description.
• Automated Telemetry: AI-generated logs and training summaries for human-readable insight into convergence status.

🔗 Model Context Protocol (MCP)

• Standardized Handshake: Custom protocol for negotiating training capabilities between Host and Edge nodes.
• Remote Validation: Validates task compatibility against a deployed MCP server on Hugging Face Spaces.
• Resource Discovery: Exposes available models and datasets via a uniform interface.

📊 Performance & Visualization

• Glassmorphism UI: Futuristic admin dashboard built with Tailwind CSS and Framer Motion.
• Real-time Analytics: Recharts implementation for tracking loss curves, accuracy, and memory usage across rounds.
• Efficient Aggregation: Weighted FedAvg algorithm with in-place tensor operations to minimize memory overhead.

Architecture

1. Frontend (React + Vite):

   • fl.worker.ts: Web Worker for background training.
   • webgpu-engine.ts: Custom WGSL compute shaders for matrix multiplication and gradient descent.
   • FLContext.tsx: State management for training lifecycle.

2. Backend (FastAPI):

   • /api/fl/updates: Receives weight deltas from browser clients.
   • /api/fl/models: Broadcasts global model weights.
   • fedavg_weighted: Aggregates updates securely.

3. MCP Server (Python):

   • Hosted on Hugging Face Spaces.
   • Provides task validation and inference capabilities.

Tech Stack

• Frontend: React 19, TypeScript, WebGPU, Vite, Tailwind CSS, Recharts
• Backend: FastAPI, PostgreSQL, SQLAlchemy
• AI/ML: Google Gemini API (Orchestration), Hugging Face Transformers, PEFT/LoRA
• DevOps: Docker Compose, Nginx

Setup & Running

Prerequisites

• Docker & Docker Compose
• Node.js 18+ (for local dev)
• Google Gemini API Key

Quick Start (Docker)

1. Clone the repository:

   git clone https://github.com/Dhanush-sai-reddy/fl-fullstack.git
   cd fl-fullstack

2. Create a .env file in frontend/:

   VITE_GEMINI_API_KEY=your_api_key_here

3. Start the stack:

   docker-compose up --build

4. Open http://localhost:5173 and select "Browser Training".

License

MIT