AI AutomationFinTech — 6 months

Predator Scalping System

Independent Project

Autonomous Trading Engine

<50ms

Execution Latency

p99 percentile

Adaptive

Regime Detection

Multi-condition

10x

Speedup

Via Numba JIT compilation

Auto

Recovery

3-state circuit breaker

The Challenge

Gold (XAU/USD) markets demand sub-50ms execution with adaptive regime detection. Manual trading cannot keep pace with micro-movements across volatile, trending, and ranging conditions.

The Solution

Built a fully autonomous scalping engine deploying 5 specialized AI agents that detect market regimes, predict micro-movements, and execute trades with institutional-grade risk management using LangGraph orchestration.

What It Does

Predator is a fully autonomous gold trading system that scalps XAU/USD by detecting market regimes (trending, ranging, volatile) and adapting its strategy in real-time. It processes tick data through 5 specialized agents, each handling a specific layer of the trading pipeline.

How It Works

The system uses adaptive regime detection (v2 algorithm) with MAD-based thresholds, Choppiness Index, and Kaufman's Efficiency Ratio. TimescaleDB stores time-series data with point-in-time correctness. A Champion/Challenger model registry enables A/B testing.

Process Flow

1Tick Data Ingestion (WebSocket/OANDA)

2Regime Classification (ADX + CHOP + ER)

3Strategy Selection (Trend/Range/Reversal)

4Position Sizing (Kelly Criterion)

5Execution (Sub-50ms via CCXT)

6Drift Monitoring (PSI/ADWIN)

Key Innovations

→Adaptive Regime Detection v2 with hysteresis filtering

→Champion/Challenger model registry with statistical testing

→Real-time drift detection (PSI, ADWIN, KS-test)

→Online/offline feature store with point-in-time correctness

→10x speedup via Numba JIT compilation

→Circuit breaker with 3-state automatic recovery

Technologies Used

Python 3.13LangGraphTimescaleDBPolarsNumPyNumbaCCXTWebSocketFastAPIStreamlitDockerRedisPostgreSQLscikit-learnADWINPSI Drift Detection

Performance Metrics

47ms

Execution Latency

End-to-end signal to execution (p99)

→ p99

88%

Regime Detection

Classification accuracy on out-of-sample data

↑ +2.3%

99.9%

System Uptime

24/7 continuous operation since deployment

→ 30d

62.5%

Strategy Win Rate

Backtested win rate on 6 months data

↑ +1.2%

View on GitHub

Agency

Interested in working together?

Let's discuss how AI enablement can transform your operations.

Get in Touch