About KARGA Markets

A technical deep-dive into AI-powered financial knowledge graphs

📊The Problem

Traditional financial data platforms store information in isolated silos. Stock prices live in one database, government contracts in another, SEC filings in a third, and prediction markets somewhere else entirely.

When you want to answer questions like "Which S&P 500 companies with significant government contracts are mentioned in prediction markets?" - you'd need to manually query multiple systems, export data, and perform complex joins in spreadsheets.

There had to be a better way.

⚡The Solution: KARGA

KARGA Markets combines three powerful technologies:

1. Knowledge Graphs

Data stored as interconnected nodes and relationships using ArangoDB, enabling complex multi-hop queries across disparate data sources in milliseconds.

2. Retrieval Augmented Generation

AI (GPT-4) generates precise database queries from natural language, then analyzes results with full context - no hallucination, only real data.

3. Semantic Search

Vector embeddings enable concept-based search - find "cybersecurity contracts" even when documents use terms like "network security" or "threat detection."

🔧System Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                          USER INTERFACE                              │
│  Next.js 14 + React + Framer Motion + Tailwind CSS                  │
│  • Natural Language Query Input                                      │
│  • Interactive Graph Visualization (ReactFlow)                       │
│  • Real-time Market Cards                                            │
│  • Data Tables with Filtering/Sorting                                │
└─────────────────────┬───────────────────────────────────────────────┘
                      │ HTTPS / REST API
┌─────────────────────▼───────────────────────────────────────────────┐
│                     FASTAPI BACKEND                                  │
│  Python 3.13 + FastAPI + Pydantic                                    │
│                                                                       │
│  ┌──────────────────────────────────────────────────────┐           │
│  │  Query Pipeline (Parallel Execution)                  │           │
│  │  ┌────────────┐  ┌──────────────┐  ┌──────────────┐ │           │
│  │  │ GPT-4      │  │ Perplexity   │  │ ArangoDB     │ │           │
│  │  │ Intent     │  │ Web Search   │  │ Graph Query  │ │           │
│  │  │ Detection  │  │ (Current     │  │ (Historical  │ │           │
│  │  │            │  │  Events)     │  │  Data)       │ │           │
│  │  └────┬───────┘  └──────┬───────┘  └──────┬───────┘ │           │
│  │       │                 │                   │         │           │
│  │       └─────────────────┴───────────────────┘         │           │
│  │                         │                              │           │
│  │                  ┌──────▼──────────┐                  │           │
│  │                  │  GPT-4 Synthesis │                  │           │
│  │                  │  Combines Results│                  │           │
│  │                  └──────────────────┘                  │           │
│  └──────────────────────────────────────────────────────┘           │
│                                                                       │
│  Security: Rate Limiting • Input Validation • HSTS Headers           │
└─────────────────────┬───────────────────────────────────────────────┘
                      │ AQL Queries
┌─────────────────────▼───────────────────────────────────────────────┐
│                    ARANGODB CLOUD                                    │
│  Multi-Model Database (Document + Graph + Search)                   │
│                                                                       │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐              │
│  │  Document    │  │  Graph       │  │  Vector      │              │
│  │  Collections │  │  Edges       │  │  Embeddings  │              │
│  └──────────────┘  └──────────────┘  └──────────────┘              │
│                                                                       │
│  • 500+ companies (S&P 500)                                          │
│  • 2M+ daily market data points (OHLCV + 40 indicators)             │
│  • 100K+ government contract awards (with embeddings)                │
│  • 50K+ SEC filings (10-K, 10-Q, 8-K)                               │
│  • 20K+ prediction markets (Polymarket + Kalshi)                    │
│  • FRED economic indicators                                          │
│  • CFTC commodity positions                                          │
└───────────────────────────────────────────────────────────────────────┘

📊Data Sources & Integration

📊 Stock Market Data

•Source: Yahoo Finance API
•Coverage: All S&P 500 companies
•Data: OHLCV, volume, market cap, P/E ratios, technical indicators (SMA, EMA, MACD, RSI, Bollinger Bands)
•Update Frequency: Daily via Airflow DAG

📋 Government Contracts

•Source: USASpending.gov API
•Coverage: Federal contract awards to public companies
•Data: Award amounts, agencies, descriptions, dates
•Special Feature: Vector embeddings for semantic search (find "AI contracts" without exact keyword match)

📈 Prediction Markets

•Sources: Polymarket API, Kalshi API
•Coverage: Politics, economics, sports, entertainment
•Data: Probabilities, volumes, liquidity, traders
•Connection: Graph edges link markets to mentioned companies (e.g., "Will Tesla reach $300?" → Tesla stock)

📋 SEC Filings

•Source: SEC EDGAR API
•Types: 10-K (annual), 10-Q (quarterly), 8-K (events)
•Processing: Parsed into sections and sentences
•Sentiment: FinBERT scores for each sentence (-1 to +1)

📊 Economic Indicators

•Source: Federal Reserve Economic Data (FRED)
•Data: S&P 500 index, Fed funds rate, unemployment, GDP, yield curves
•Coverage: Historical time series data

📊 Commodity Positions

•Source: CFTC Commitments of Traders Report
•Data: Long/short positions by trader type (commercial, non-commercial, retail)
•Commodities: Oil, gold, wheat, corn, natural gas, etc.

🔗Knowledge Graph Structure

Data isn't just stored—it's connected. Here's how relationships enable powerful queries:

Company → Market Data

HAS_MARKETDATA edges connect companies to their daily stock prices, enabling queries like "Show me tech companies with SMA_50 > SMA_200 (golden cross)"

Company → Government Awards

HAS_AWARD edges link companies to contracts, enabling semantic searches: "Defense companies with cybersecurity contracts over $10M"

Market → Company

market_mentions_company edges connect prediction markets to mentioned tickers: "Tesla reaches $300" → TSLA

Company → SEC Filings → Sentences

HAS_FILING → has_section → has_sentenceMulti-hop traversal for sentiment analysis: "Show negative FinBERT sentences from recent Apple 10-Ks"

Company → Commodity Positions

HAS_COMMODITY_POSITION links companies to CFTC data for commodity exposure analysis

💡 Example Multi-Hop Query

"Find energy companies with government contracts mentioning 'renewable' that are mentioned in prediction markets with volume > $50k"

→ Traverses Company → Awards (semantic search) → Markets (graph join) in milliseconds

🤖AI Query Pipeline

When you ask a question, here's what happens behind the scenes:

Step 1: Intent Detection

GPT-4 classifies your query: Is it about a specific ticker (AAPL, MSFT) or a concept (AI, cybersecurity)? This determines whether to use exact matching or semantic search.

Input: "Show me AI companies with government contracts"
Intent: concept_query

Step 2: Query Planning

GPT-4 receives the full database schema (collections, fields, relationships) and generates optimized AQL (ArangoDB Query Language) with proper joins and filters.

FOR award IN Award
  FILTER COSINE_SIMILARITY(
    award.description_embedding,
    @query_vector
  ) >= 0.75
  FOR company IN Company
    FILTER company.ticker == award.ticker
    RETURN {company, award}

Step 3: Parallel Execution

Two queries run simultaneously:

• Database Query: AQL executes against ArangoDB (historical data)
• Web Search: Perplexity searches for current events (real-time context)

Step 4: Synthesis & Analysis

GPT-4 combines database results with web context, analyzes patterns, and generates:

• Markdown Tables: Formatted results with key metrics
• Insights: Trends, correlations, anomalies
• Follow-up Questions: Suggested deeper dives

🔧Technology Stack

Frontend

Framework: Next.js 14 (App Router)
UI: React 18, TypeScript
Styling: Tailwind CSS
Animations: Framer Motion
Graph Viz: ReactFlow
Hosting: Vercel

Backend

Framework: FastAPI (Python 3.13)
Validation: Pydantic
Security: SlowAPI rate limiting
LLM: OpenAI GPT-4
Web Search: Perplexity AI
Hosting: Railway

Database

Platform: ArangoDB Cloud
Type: Multi-model (Document + Graph)
Query Language: AQL
Embeddings: OpenAI text-embedding-3-small
Size: ~5GB (2M+ documents)
Location: Germany (GDPR compliant)

Data Pipeline

Orchestration: Apache Airflow
Processing: Python, Pandas, NumPy
Sentiment: FinBERT
Schedule: Daily updates at 2 AM UTC
Monitoring: Airflow UI + logs

⚡Performance & Scale

< 3s

Average Query Time

(DB + AI analysis)

2M+

Documents in Graph

(Companies, markets, filings)

50ms

Graph Traversal Time

(3-hop relationships)

Performance Optimizations

✓Persistent indexes on ticker, date, volume fields
✓Skip-list indexes for range queries
✓Edge collections for O(1) relationship lookups
✓Query result caching (5-minute TTL)
✓Parallel DB + web search execution
✓Streaming results with batch_size=1000

📈Future Enhancements

⚡ Real-time Data

WebSocket connections for live market data updates, streaming prediction market probability changes as they happen.

📊 Portfolio Tracking

User accounts to track favorite companies, save queries, and set up alerts for specific market conditions.

🤖 Advanced ML Models

Time-series forecasting with LSTM, anomaly detection for unusual trading patterns, correlation discovery between data sources.

📊 More Data Sources

Twitter sentiment, Reddit discussions, earnings call transcripts, patent filings, and international market data.

📊 Custom Dashboards

Drag-and-drop dashboard builder with custom charts, metrics, and KPIs tailored to individual research needs.

🔗 API Access

Public API with authentication for programmatic access to KARGA capabilities, enabling integrations with trading platforms.

🔓Open Source & Contributions

KARGA Markets is built with transparency in mind. While the core application is proprietary, we're exploring open-sourcing components of the query planning system and graph schema to help others build similar systems.

Interested in collaborating or have ideas for improvement? Reach out at karga.analytics@gmail.com

← Back to Platform