Crypto AI Agent

March 25, 2026

Overview

An AI-powered crypto assistant that interprets user queries and delivers structured responses using a routed LangGraph architecture, specialized tools, and structured data pipelines.

The system is designed to produce reliable, context-aware outputs while remaining resource-efficient and scalable.

Architecture

The system follows a structured pipeline:

• Flask backend handles incoming requests
• LangGraph agent routes queries based on intent
• Specialized tools handle data retrieval and analysis
• Frontend renders results and visualizations

The agent dynamically selects execution paths instead of relying on a single generalized prompt.

 Query Routing

User queries are often unstructured and ambiguous.

A lightweight classification layer determines:

• Whether the query is crypto-relevant
• Whether a specific asset is referenced
• The appropriate category and execution path

This routing ensures each query is processed with the correct tools and prompts, improving both efficiency and response quality.

Tooling & Data Strategy

The system follows a tool-first design, minimizing reliance on raw LLM reasoning.

Market Data

• Prices are stored locally and updated via scheduled API calls
• External API usage is minimized (daily updates per asset)
• Reduces latency, cost, and dependency on real-time calls

Technical Analysis

• Indicators are precomputed from structured price data
• The LLM receives clean, formatted inputs instead of raw data
• Prompts explicitly define how each indicator should be interpreted

News

• Relevant articles are selected from trusted sources
• The system prioritizes high-signal, popular content over volume

FAQ Handling

• Common crypto questions are handled via predefined prompt-based responses
• Ensures consistent answers for foundational queries

Visualization Layer

• Charts are generated on the client side using Plotly (JavaScript)
• Backend passes structured JSON instead of retaining chart objects
• This reduces memory usage and improves performance

Constraints & Optimization

A 512MB RAM constraint was used in production to evaluate system efficiency.

This exposed performance bottlenecks and led to major optimizations:

• Refactoring of over 2,000 lines of code
• Removal of unnecessary object retention
• Consolidation of logic where appropriate
• Transition to a stateless (one-shot) agent design
• Migration of chart rendering to the frontend

These changes significantly improved stability and resource usage.

Key Design Decisions

• Routed agent architecture using LangGraph
• Tool-first approach instead of LLM-centric logic
• Structured inputs for deterministic outputs
• Stateless query processing model
• Client-side rendering for visualization efficiency

Outcomes

• Efficient handling of ambiguous, real-world user queries
• Reduced API usage and operational costs
• Improved system stability under constrained resources
• Modular architecture that separates routing, tools, and presentation

Resources

• Live Demo: https://aishepherd.dev/portfolio/crypto-agent
• Walkthrough Video: https://www.youtube.com/watch?v=dT7XhyRpzNk
• Code Repository: https://github.com/The-AI-Shepherd/Crypto_AI_Agent