🚀 StockSense Agent – Multi-Agent Financial Intelligence System

StockSense Agent is a full-stack Generative AI system that analyzes real-time news and explains its potential impact on stock behavior using a multi-agent architecture.

⚠️ This system provides AI-driven insights, not financial advice or stock price predictions.

---

🌐 Live Demo

🔗 Live Link: Live Link

---

🖼️ Screenshots

📊 Application Interface

🧠 Analysis Output

---

🧠 System Overview

StockSense Agent is designed as a multi-agent pipeline where each component performs a specific responsibility:

User Query
   ↓
Router Agent (intent + entity extraction)
   ↓
Validation Layer (rule-based + LLM)
   ↓
Orchestrator
   ↓
 ┌───────────────┬───────────────┐
 ↓               ↓               ↓
News Agent    Finance Agent   Memory
 ↓               ↓
      Reasoning Agent (LLM)
               ↓
        Structured Output
               ↓
           Frontend UI

---

🤖 Agents Description

1. 🧭 Router Agent

• Extracts intent (news, impact, general)
• Identifies company names using LLM + fallback
• Handles vague and multi-company queries

---

2. 📰 News Agent

• Fetches real-time news using Tavily API
• Filters and structures relevant company news
• Uses caching to reduce API calls

---

3. 📊 Finance Agent

• Resolves company → ticker
• Fetches stock-related data
• Provides structured financial context

---

4. 🧠 Reasoning Agent

• Uses LLM to:

  • Analyze news impact
  • Generate reasoning chain
  • Identify risks
  • Estimate confidence

• Adds deterministic conclusion layer

---

5. 🛡️ Validation Layer

• Prevents invalid or random inputs

• Hybrid approach:

  • Rule-based filtering
  • LLM-based verification

---

6. 🧠 Memory Module

• Stores:

  • last company
  • last query

• Enables multi-turn conversations

---

7. ⚡ Caching Layer

• Reduces latency and API cost

• TTL-based caching for:

  • news data
  • financial data

---

🔄 Data Flow (Step-by-Step)

1. User submits a query via frontend
2. Router Agent extracts intent and company
3. Validation layer filters invalid inputs
4. Orchestrator controls pipeline execution
5. News + Finance agents fetch data
6. Reasoning agent analyzes impact
7. Structured response is returned
8. Frontend renders insights

---

📡 API Endpoints

🔹 POST /analyze

Analyze a company based on real-time news

Request:

{
  "query": "Tesla latest news impact",
  "user_id": "user123"
}

Response:

[
  {
    "company": "Tesla",
    "event_summary": ["News headline..."],
    "event_type": "earnings / macro",
    "impact_direction": "Positive",
    "reasoning_chain": ["Step 1...", "Step 2..."],
    "risks": ["Risk 1...", "Risk 2..."],
    "confidence": "medium",
    "conclusion": "Stock may increase",
    "disclaimer": "This is an AI-generated analysis..."
  }
]

---

🔹 Error Response

{
  "message": "Please enter a valid company name (e.g., Tesla, Apple)"
}

---

⚙️ Tech Stack

• Backend: FastAPI
• Frontend: Streamlit
• LLM: OpenAI OSS / Groq
• Search: Tavily API
• Architecture: Multi-Agent System
• Deployment: Render and streamlit cloud

---

⚠️ Limitations

• LLM responses may vary
• Not suitable for real financial decision-making
• Dependent on external APIs

---

📌 Key Engineering Highlights

• Multi-agent orchestration instead of single LLM call
• Hybrid validation to reduce hallucination
• Deterministic + LLM combined reasoning
• Session-based conversational memory
• Production-style backend/frontend separation

---