Local Agentic System - Oxford AI Summit 2025

Enterprise-Ready Multi-Agent AI System (100% Local Execution)

Overview

This repository contains a modular, production-grade multi-agent AI framework that runs entirely on local hardware. It orchestrates multiple specialized agents for advanced reasoning, research, validation, and quality assurance—no cloud required.

Features

• 100% local execution (no external dependencies)
• Multi-agent orchestration with streaming responses
• Modular, extensible agent framework
• Real-time web interface
• Runs on consumer GPUs or Apple Silicon

Quick Start

Prerequisites

• Python 3.10+
• Ollama installed
• 18GB+ VRAM recommended (or Apple Silicon with 32GB+ memory)

Installation

# Clone the repository
git clone https://github.com/Magnussmari/Oxford-ai-summit-2025-Local-Agents.git
cd Oxford-ai-summit-2025-Local-Agents

# Install dependencies
pip install -r demo/requirements.txt

# Download required models
./models/pull_models.sh

# Launch the demo
cd demo && ./launch.sh

Visit http://localhost:8000 to use the system.

Agent Team

Agent	Model	Purpose
Principal Synthesizer	deepseek-r1:8b	Reasoning & orchestration
Domain Specialist	qwen3:8b	Expert domain analysis
Web Harvester	qwen3:4b	Web research
Fact Validator	phi4-mini	Claim verification
Quality Auditor	phi4-mini	Output assessment

Repository Structure

demo/
  app.py           # FastAPI server
  agents/          # Agent implementations
  static/          # Web UI (HTML/CSS/JS)
  utils/           # Monitoring utilities
  launch.sh        # Demo launcher
models/
  pull_models.sh   # Download models

Customization Guide

Adding a New Agent

1. Create your agent in demo/agents/your_agent.py:

from .base import PresentationAgent
class YourAgent(PresentationAgent):
    def __init__(self):
        super().__init__(name="Your Agent", model="model_name:tag", role="specialist", temperature=0.7)
    async def process(self, query: str, context: dict, stream_callback=None) -> str:
        prompt = f"Your prompt: {query}"
        return await self.run(prompt, stream_callback)

2. Register it in agents/orchestrator_enhanced.py.

Modifying Agent Behavior

• Change models: edit agent __init__
• Adjust prompts: modify agent prompt templates
• Orchestration: update orchestrator_enhanced.py
• UI: edit files in static/

FAQ

Q: Can I use different models?
A: Yes, any Ollama-compatible model. Update the agent class.

Q: How do I enable web search?
A: Add a Brave Search API key to .env as BRAVE_API_KEY.

Q: What are the hardware requirements?
A: 18GB VRAM recommended, but smaller models can run on 8-12GB GPUs.

Troubleshooting

• Models not found: Run ./models/pull_models.sh
• Out of memory: Use fewer agents or smaller models
• WebSocket errors: Ensure port 8000 is open
• Import errors: Run commands from the project root

License

MIT License. See LICENSE for details.