Multi-agent Marketing Automation System (NTI Final Project)

Overview

Multi-agent Marketing Automation System is an advanced AI-powered solution that enhances marketing efficiency through intelligent automation. It prioritizes fine-tuning with DeepSeek-R1-Distill-Llama-8B for chatbot interactions, Vision Transformer (ViT) for real-time image deduplication, and a multi-agent system for automating market analysis, content creation, and campaign optimization.

Demo: Click here

---

Features

1. Assistant Chatbot

• Uses Retrieval-Augmented Generation (RAG) for real-time, accurate responses to frequently asked questions (FAQs).
• Supports both English and Arabic.

2. Real-Time Image Deduplication System

• Utilizes Vision Transformer (ViT) to detect and eliminate near-duplicate images in real time.
• Enhances media asset management by reducing redundant images.

3. Multi-Agent Marketing System

This system automates key marketing tasks through specialized AI agents:

🏆 Market Analysis Agent

• Tracks market trends and competitor activity.
• Automates report generation and email campaign execution.

✍️ Content Marketing Agent

• Automates the creation of social media posts and blog content.
• Ensures consistent and engaging brand communication.

📈 Campaign Optimization Agent

• Uses Agglomerative Clustering for customer segmentation.
• Optimizes targeted email campaigns for higher conversion rates.

📊 Data Analysis Agent

• Performs data preprocessing before generating insights.
• Chooses the most appropriate chart type to create interactive dashboards.
• Helps businesses make data-driven marketing decisions.

---

Fine-Tuning DeepSeek-R1 for Telecom Chatbots

1. Instruction Fine-Tuning

• Transformed a telecom dataset into an instruction dataset.
• Fine-tuned DeepSeek-R1-Distill-Llama-8B using QLoRA.
• Deployed the fine-tuned model on Hugging Face Hub.
• Evaluated chatbot performance using llama-3.3-70B-Instruct.
• Hugging Face Model: moo100/DeepSeek-R1-telecom-chatbot

2. Preference Fine-Tuning

• Created a preference dataset with Phi4, generating preferred responses via evaluation prompts.
• Fine-tuned DeepSeek-R1-Telecom-Chatbot using Unsloth and DPOTrainer.
• Hugging Face Model: moo100/DeepSeek-R1-telecom-chatbot-v2

---

Contributors

This project was developed as part of the NTI Final Project by a team of AI and marketing automation enthusiasts:

• Mohamed Abdulaziz
• Mahmoud Elwaled
• Mariam Mustafa
• Amr Abdelatey
• Ahmed Ali

---