TripleGain

Multi-Asset LLM-Assisted Trading System

An autonomous cryptocurrency trading system using a multi-agent architecture with 6 LLMs for decision comparison and consensus-based trading.

Overview

Aspect	Details
Objective	Grow BTC, USDT, XRP holdings autonomously
Target Allocation	33% BTC / 33% XRP / 33% USDT
Trading Pairs	BTC/USDT, XRP/USDT, XRP/BTC
Starting Capital	~$2,100
Exchange	Kraken (spot + margin)
LLMs	Qwen(local), GPT, Grok, Claude, DeepSeek
Deployment	Paper trading → Micro-live → Scale

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                        ANALYSIS LAYER                             │
│  Technical Analysis │ Regime Detection │ Sentiment Analysis       │
│     (Qwen Local)    │   (Qwen Local)   │    (Grok + GPT)          │
└──────────────────────────────┬───────────────────────────────────┘
                               ▼
┌──────────────────────────────────────────────────────────────────┐
│                    DECISION LAYER (6-Model A/B)                   │
│  GPT │ Grok │ DeepSeek V3 │ Claude Sonnet │ Claude Opus │ Qwen   │
└──────────────────────────────┬───────────────────────────────────┘
                               ▼
┌──────────────────────────────────────────────────────────────────┐
│  Risk Management (Rules) ──► Coordinator ──► Order Execution     │
│     VETO AUTHORITY              │                                │
└──────────────────────────────────────────────────────────────────┘

Key Design Decisions:

• 6-Model A/B Testing: All LLMs run in parallel for comparison and consensus
• Rules-Based Risk: Deterministic risk management, no LLM override possible
• Trend-Following: Research shows mean reversion fails on crypto
• Conservative Execution: Quality over quantity

Quick Start

# Start database
docker-compose up -d timescaledb

# Fill any data gaps
python -m data.kraken_db.gap_filler --db-url "$DATABASE_URL"

# Run tests
pytest

Project Status

Current Phase: Pre-Phase 1 (Infrastructure Ready)

Phase	Status	Description
Infrastructure	Complete	TimescaleDB, data collectors, Ollama
1. Foundation	Not Started	Indicators, snapshots, prompts
2. Core Agents	Not Started	TA, Regime, Risk, Trading Decision
3. Orchestration	Not Started	Communication, Coordinator, Execution
4. Extended	Not Started	Sentiment, Hodl Bag, Dashboard
5. Production	Not Started	Testing, Paper/Live Trading

Data Infrastructure

• Historical Data: 5-9 years via TimescaleDB continuous aggregates
• Symbols: XRP/BTC (2016), BTC/USDT (2019), XRP/USDT (2020)
• Timeframes: 1m, 5m, 15m, 30m, 1h, 4h, 12h, 1d, 1w
• Collectors: WebSocket writer, gap filler, order book, private trades

Target Metrics

Metric	Target
Annual Return	> 50%
Maximum Drawdown	< 20%
Sharpe Ratio	> 1.5
Win Rate	> 50%
System Uptime	> 99%

Risk Controls

• Max Leverage: 5x
• Daily Loss Limit: 5%
• Weekly Loss Limit: 10%
• Max Drawdown Circuit Breaker: 20%
• Required Stop-Loss on all trades

Documentation

Document	Description
Master Design	Complete system design
Implementation Plan	5-phase roadmap
Multi-Agent Architecture	Agent specifications
Risk Management	Risk rules engine
Kraken API Reference	Exchange integration

Tech Stack

• Language: Python 3.11+
• Database: TimescaleDB (PostgreSQL extension)
• LLM Local: Ollama (Qwen 2.5 7B)
• LLM API: OpenAI, Anthropic, xAI, DeepSeek
• Dashboard: React (planned)
• Exchange: Kraken REST/WebSocket API