AutoStream AI Agent

LIVE DEMO: https://autostream-ai-agent-5yfpqwgetdjlaffptrajvp.streamlit.app

This is a Conversational AI Agent built for AutoStream, a fictional SaaS company. The agent is designed to qualify sales leads, answer pricing questions using RAG, and capture user details (Name, Email, Platform) into a structured format.

🛠️ Tech Stack

• Framework: LangGraph (Stateful Multi-Turn Conversations)
• LLM: Google Gemini 2.5 Flash
• Knowledge Base: JSON-based RAG
• Logic: Hybrid Approach (LLM Context Extraction + Persistent State)

🚀 Setup Instructions

1. Clone the Repository

   git clone [https://github.com/Ujjwal-Sharma-Data/AutoStream-AI-Agent.git](https://github.com/Ujjwal-Sharma-Data/AutoStream-AI-Agent.git)
   cd AutoStream-AI-Agent

2. Install Dependencies

   pip install -r requirements.txt
   

3. Configure API key
   • Create a .env file in the root folder.
   • Add your Google Gemini API key:

     GOOGLE_API_KEY="your_actual_key_here"
     

4. Run the Agent

   python3 main.py
   

Architecture Explanation

The system follows a cyclic State Machine architecture:

1. Input: User message enters the graph and is appended to the persistent conversation history.
2. Analysis (The Brain): The agent_node constructs a prompt containing the full transcript and the current known state. The LLM analyzes this to extract new entities (Name, Email, Platform) contextually.
3. State Merge: Extracted entities are merged into the persistent lead_data dictionary.
4. Decision:
   • Missing Data: The agent asks a follow-up question.
   • Complete Data: The tool_check_node validates the state and executes mock_lead_capture.
5. Output: The final response is returned to the user.

Technical Decisions (Why LangGraph?)

I selected LangGraph over AutoGen because this task requires a deterministic control flow. Unlike AutoGen (which is better for open-ended multi-agent debates), LangGraph allows us to define strict loops—ensuring the agent cannot proceed to the "Tool Trigger" step until specific data criteria (Name/Email/Platform) are met.

State Management: We utilize a Hybrid Persistence model. Instead of relying solely on the LLM's short-term memory, we maintain a structured AgentState dictionary. Every turn, the LLM acts as an extractor, parsing the transcript to find new information which is permanently saved to the lead_data state. This ensures "perfect memory" even if the user provides details out of order or changes topics.

WhatsApp Deployment Plan

To deploy this agent on WhatsApp, I would integrate the WhatsApp Business API with a Python webhook:

• Webhook: A FastAPI endpoint receives incoming HTTP POST requests from Meta containing the user's message and wa_id (phone number).
• Session Handling: The wa_id serves as the unique session key. We retrieve the user's persisted LangGraph state from a Redis database using this key.
• Execution: The message passes through the graph, updates the state, and the agent's text response is sent back via the messages API endpoint.