OrderGuard: Safe AI Trading

OrderGuard is a multi-agent AI trading research and paper-trading system built around one core principle: AI can make recommendations, but it cannot bypass deterministic safety controls. The system starts with Alpaca market data and uses a Market Intelligence Agent to identify the current market regime. A Discovery Agent then proposes quantitative trading strategies, which are evaluated through chronological backtesting and out-of-sample validation. Promising strategies are challenged by an Adversary Agent that stress-tests parameter sensitivity, concentration, and regime dependence before assigning a robustness score and PASS/WATCH/REJECT verdict. A strategy lifecycle engine then determines whether a strategy remains ALIVE, enters WATCH, or is KILLED. The Portfolio Manager allocates capital among surviving strategies while maintaining safety constraints. Before any trade reaches Alpaca, it must pass OrderGuard's deterministic risk layer, including position limits, buying power, drawdown protection, strategy status, and market-session checks. These controls are implemented independently of the LLM and cannot be overridden by it. Approved trades can then be sent to Alpaca's paper-trading environment, while the Performance Monitor tracks strategy performance and can reduce allocation or terminate deteriorating strategies. The project also includes automated testing, GitHub Actions CI/CD, Docker containerization, and GitHub-based deployment infrastructure. OrderGuard is designed not as an AI that blindly trades, but as an AI system that can research, challenge, and explain trading decisions while keeping final execution authority behind deterministic safety controls.