Value Investing Multi-Agent Research System

A self-hosted multi-agent AI system that discovers and researches special situation investing opportunities in Indian markets (NSE/BSE).

Features

• Autonomous Discovery: Monitors NSE/BSE filings daily for insider trading, corporate announcements, and board meetings.
• Deep Research: 4-level investigation agents (Context, History, Fundamentals, Synthesis) driven by DeepSeek V3.
• Persistent Memory: Uses PostgreSQL with pgvector for semantic search and Redis for caching.
• Learning System: Feedback loop trains a Reward Model to personalize recommendations over time.
• Local Deployment: Fully containerized with Docker Compose.

Architecture

The system uses LangGraph to orchestrate a multi-agent research pipeline with conditional routing and state management.

graph TD
    Start([Daily Scheduler]) --> Scraper[Data Collector<br/>NSE/BSE Filings]
    Scraper --> Signals[(Raw Signals<br/>Insider Trades, Announcements)]
    
    Signals --> Discovery[Discovery Agent<br/>Qwen-2.5-72B<br/>Fast Filter]
    
    Discovery --> Gate{Is Interesting?}
    Gate -->|No| End1([Discard])
    Gate -->|Yes| DeepResearch[Deep Research Agent<br/>DeepSeek V3<br/>4-Level Analysis]
    
    DeepResearch --> Level1[Level 1: Company Basics]
    Level1 --> Level2[Level 2: Historical Patterns]
    Level2 --> Level3[Level 3: Fundamentals]
    Level3 --> Level4[Level 4: Synthesis]
    
    Level4 --> Context[Context Agent<br/>DeepSeek V3<br/>Industry & Peers]
    
    Context --> Validation[Validation Agent<br/>Qwen-2.5-72B<br/>Fact Checking]
    
    Validation --> Synthesis[Synthesis Agent<br/>DeepSeek V3<br/>Final Report]
    
    Synthesis --> QualityGate{Score >= 7.0?}
    QualityGate -->|No| End2([Discard])
    QualityGate -->|Yes| Store[(PostgreSQL<br/>+ pgvector)]
    
    Store --> Digest[Daily Digest<br/>HTML Report]
    
    style Discovery fill:#e1f5ff
    style DeepResearch fill:#fff4e1
    style Context fill:#fff4e1
    style Validation fill:#e1f5ff
    style Synthesis fill:#fff4e1
    style Store fill:#e8f5e9

Loading

Key Components

1. Data Collector: Scrapes NSE and Screener.in for signals
2. Discovery Agent (Qwen-2.5-72B): Fast, cost-effective filtering
3. Deep Research Agent (DeepSeek V3): Multi-level comprehensive analysis
4. Context Agent (DeepSeek V3): Industry trends and peer comparison
5. Validation Agent (Qwen-2.5-72B): Fact-checking and consistency validation
6. Synthesis Agent (DeepSeek V3): Generates structured investment memos
7. Quality Filter: Only insights scoring ≥7.0 are saved

Prerequisites

• Python 3.11+
• Docker & Docker Compose
• API Keys:
  • OpenRouter API Key: Single key to access both DeepSeek (Research) and Qwen (Discovery) models.

Installation

1. Clone the repository:

   git clone <repo_url>
   cd value-investing-agents

2. Configure Environment: Copy .env.example to .env (or just edit .env if created):

   # Edit .env with your API keys
   nano .env

3. Start Infrastructure:

   docker-compose up -d

4. Install Python Dependencies:

   pip install -r requirements.txt

   Note: You may need to install Playwright browsers:

   playwright install

5. Initialize Database: The database should initialize automatically via docker-compose, but if needed:

   psql -h localhost -U researcher -d value_investing_research -f scripts/init_db.sql

Usage

Run Daily Research

Manually trigger the daily research pipeline:

python3 scripts/daily_run.py

This will collect signals, run agents, and generate a daily digest in /tmp/daily_digest_YYYYMMDD.html.

Note: The first run will download ~1.5GB of embedding models. This may take a few minutes depending on your internet connection.

Rate Insights

Start the feedback UI to rate generated insights and train the model:

python3 -m src.feedback.collector

Open http://localhost:5000 in your browser.

Automated Scheduling

To run the scheduler in the background:

python3 scripts/scheduler.py

Project Structure

value-investing-agents/
├── docker-compose.yml      # Infrastructure
├── .env                    # Config
├── src/
│   ├── agents/             # Agent logic (Discovery, Research, etc.)
│   ├── graph/              # LangGraph workflow
│   ├── memory/             # Postgres & Redis managers
│   ├── models/             # LLM & Embedding clients
│   ├── scrapers/           # NSE & Screener scrapers
│   └── feedback/           # Flask app
├── scripts/                # Execution scripts
└── data/                   # Docker volumes