Agentic Crypto Trading System

Architecture Components

The system is designed with a modern, decoupled architecture ensuring high scalability and rapid AI execution.

Frontend

Built with Angular 21, providing a UI to view portfolio status, recent trades, logs, and trigger new trading cycles manually.

Backend

Developed using Python and FastAPI, handling REST APIs, database connections, and triggering AI workflows.

Database

MongoDB is used for persistent storage of portfolios, trade histories, and execution logs, with an in-memory fallback.

LangGraph

Models core intelligence as a state graph, allowing structured data flow from data gatherers to execution managers.

Core Intelligence

AI Integration

Integrates Google's Gemini-3.1-pro model to process complex data and output structured trading decisions and rigorous risk reasoning.

Agentic Workflow

The core trading intelligence is powered by a multi-agent system orchestrated by LangGraph. Each agent plays a specific role.

1 Market Data Gatherer

Role: Collects real-time market data (price, volume, trend).

Value: Provides foundational quantitative context.

2 Sentiment Analyst

Role: Analyzes news and social media to provide a sentiment score.

Value: Captures emotional and qualitative market drivers.

3 Technical Analyst

Role: Computes technical indicators like RSI and MACD.

Value: Identifies short-term momentum and price patterns.

4 Strategy/Decision Agent

Role: Aggregates data and uses Gemini to formulate a strategy (ACTION, AMOUNT, REASONING).

Value: Acts as the brain, synthesizing diverse streams to make context-aware decisions.

5 Risk Manager (LLM)

Role: Evaluates proposed trades against standard risk management principles to APPROVE or REJECT.

Value: Protects portfolio from severe drawdowns, ensuring long-term capital preservation.

6 Execution Agent

Role: Executes the order if approved by Risk Management.

Value: Automates physical trade execution.

Control & Communication Flow

Agents do not communicate directly. A shared AgentState object is managed by the LangGraph state machine. As the graph executes sequentially, each agent receives the state, performs its task, appends findings, and passes the enriched state down the pipeline.

System Diagrams

Overall System Design

graph TD
    User([User]) -->|Interacts with UI| Frontend[Angular Frontend]
    Frontend <-->|REST API| Backend[FastAPI Backend]

    subgraph Server Infrastructure
        Backend -->|Persists Data| MongoDB[(MongoDB Database)]
        Backend -->|Invokes Workflow| LangGraph[LangGraph Engine]
    end

    subgraph AI Orchestration
        LangGraph <-->|Prompts & Responses| LLM[Google Gemini LLM API]
    end

LangGraph State Flow

graph TD
    Start((Start)) --> gather_market_data

    gather_market_data -->|"Market Data Gathered, Sentiment and Technicals Analyzed"| formulate_strategy
    formulate_strategy -->|"Strategy Proposed (via LLM)"| evaluate_risk
    evaluate_risk -->|"Risk Evaluated (via LLM)"| execute_trade

    execute_trade -->|"Trade Executed (or Hold)"| End((End))

    %% Notes attached via dotted lines
    note_strategy["Uses LLM to decide
BUY, SELL, or HOLD based on
aggregated data."]
    formulate_strategy -.- note_strategy

    note_risk["Uses LLM to rigorously
evaluate the proposed strategy
against risk principles."]
    evaluate_risk -.- note_risk

    %% Styling to visually distinguish the notes
    classDef note fill:#fff5ad,stroke:#d4c257,stroke-width:1px,color:#333;
    class note_strategy,note_risk note;

Setup & Run

Get the agentic trader running locally in minutes.

Backend Setup

1

Navigate to backend
cd backend
2

Install packages
pip install -r requirements.txt
3

Set API Key
export GOOGLE_API_KEY="..."
4

Start server
uvicorn main:app --reload

Frontend Setup

1

Navigate to frontend
cd frontend
2

Install packages
npm install
3

Start UI
npm start