Corporate Mail Assistant 🤖📧

An intelligent AI agent that automatically reads, classifies, and drafts replies to incoming corporate emails. Built with Python, LangGraph, and Hugging Face Transformers (BERT & Gemma).

Workflow

Hugging Face Model

You can find and use the model on the Hugging Face Hub here:

darshandugar/MailClassifier-DistilBERT & darshandugar/Corporate-Email-Response-Generator-gemma-2b

Features

• Automated Email Polling: Continuously monitors a Gmail inbox for new, unread emails.
• Intent Classification: Uses a fine-tuned BERT model to classify the intent of each email (e.g., "Merger Announcement," "Sustainability Initiative").
• Dynamic Workflow: Leverages LangGraph to route tasks based on the classified intent.
• Contextual Reply Generation: Uses a fine-tuned Gemma-2B model (with LoRA adapters) to generate context-aware, professional draft replies.
• End-to-End Automation: Automatically sends the generated reply and marks the original email as read.

How It Works: The Workflow

This project uses LangGraph to create a stateful graph that mimics a decision-making process.

1. The main loop in recieve_mail.py polls the Gmail inbox.
2. When an unread email is found, its content is passed to the run_workflow function from main_graph.py.
3. LangGraph Workflow Begins:
   • Node 1: classify_intent: The email content is processed by the IntentClassifier (fine-tuned BERT) to determine its purpose.
   • Conditional Edge: Based on the intent, the graph routes to a specific handler node (e.g., handle_merger).
   • Node 2: handle_...: The handler node processes the request, gathers necessary details (e.g., pulling RAG data, creating a ticket ID).
   • Node 3: generate_response: The intent and details are passed to the EmailGenerator (fine-tuned Gemma-2B), which drafts a complete, formal reply.
4. The final reply (subject and body) is returned to recieve_mail.py.
5. The script uses the Gmail API to send the generated reply and marks the original email as read.

Project Structure

.
├── .gitignore               # Tells Git what to ignore
├── config.py                # Central config for model paths
├── intent_classify.py       # Class for the BERT intent classifier
├── main_graph.py            # Defines the core LangGraph workflow
├── recieve_mail.py          # Main script: handles Gmail auth, polling, & sending
├── reply_generator.py       # Class for the Gemma reply generator
├── requirements.txt         # List of Python dependencies
│
├── credentials.json         # YOUR Google API credentials (Must add)
├── token.json               # Generated by Google Auth (Ignore)
│
├── mail_category/           # YOUR fine-tuned BERT model files
│   ├── config.json
│   ├── pytorch_model.bin
│   └── ...
│
└── fine_tuned_gemma_2b_adapters/ # YOUR fine-tuned Gemma adapter files
    ├── adapter_config.json
    ├── adapter_model.safetensors
    └── ...

Setup & Installation

1. Clone the Repository

git clone https://github.com/YourUsername/Corporate-Mail-Assistant.git
cd Corporate-Mail-Assistant

2. Set up Google API Credentials

1. Go to the Google Cloud Console.
2. Create a new project.
3. Enable the Gmail API.
4. Create credentials for an "OAuth client ID".
5. Select "Desktop app" as the application type.
6. Download the JSON credentials file and save it in the project root as credentials.json.

3. Add Your Fine-Tuned Models

This project assumes you have already fine-tuned your models.

1. BERT Intent Model: Place your fine-tuned BERT model files into the mail_category folder.
2. Gemma Adapters: Place your fine-tuned Gemma LoRA adapters into the fine_tuned_gemma_2b_adapters folder.

4. Install Dependencies

Install all the required Python packages:

pip install -r requirements.txt

How to Run

1. First-time Authentication:

   • Run the main script from your terminal:

     python recieve_mail.py

   • Your browser will automatically open, asking you to authorize the application to access your Gmail account.
   • Grant permissions. The script will create a token.json file to store your credentials for future runs.

2. Run the Assistant:

   • Once authenticated, the script will start running:

     --- Configuration Loaded (Local Paths) ---
     ...
     --- LANGGRAPH WORKFLOW COMPILED ---
     --- LOADING MODELS (This happens once) ---
     ...
     --- MODELS LOADED ---
     Starting mail attender service...
     Service running. Ignoring emails received before startup.
     Checking for new mail every 30 seconds...
     

   • The assistant is now active. When a new email arrives in the authorized inbox, it will process it automatically.