AI-Based Real-Time Decision Systems

Abstract

Real-time AI decision systems -- systems that must perceive, reason, and act within strict latency constraints in dynamic environments -- represent one of the most demanding classes of applied AI. Millisecond-scale decisions in financial trading, sub-second responses in industrial control, second-scale triage in emergency dispatch, and near-real-time recommendations in dynamic pricing each impose different latency, accuracy, and safety requirements that fundamentally shape system architecture. This study evaluates nine real-time AI decision systems across six application domains: algorithmic trading, industrial process control, emergency medical dispatch triage, dynamic content recommendation, autonomous vehicle intersection management, and cybersecurity intrusion response. Systems evaluated include lightweight streaming ML (River, online XGBoost), sub-millisecond neural inference (TensorRT-optimised CNNs), transformer-based real-time NLP (RoBERTa-RT), reinforcement learning controllers (PPO-RT), event-driven graph processing (GraphSAGE-Stream), LLM-assisted triage (TriageLLM), multi-modal fusion (MMRT), and an adaptive latency-accuracy framework (ALAF). Evaluation covers decision latency (mean and P99), accuracy under latency constraints, graceful degradation under system load, and safety validation for safety-critical domains. TensorRT-optimised CNN achieves P99 latency of 0.84ms for industrial vision inspection. PPO-RT reduces intersection congestion by 22.4% vs. rule-based traffic control. TriageLLM achieves 84.2% agreement with senior paramedic triage decisions at 4.2-second mean latency. ALAF dynamically balances latency and accuracy achieving the best composite score across domains. A latency-accuracy-safety framework for real-time AI system design is proposed