AI-Driven Self-Healing High-Frequency Trading (HFT) Infrastructure for Fault Prediction and Latency Optimization

High-Frequency Trading (HFT) systems operate under extreme latency requirements, where microsecond deviations can result in significant financial impact. This paper introduces a cloud-native, agent-based self-healing infrastructure utilizing large language models (LLMs) for fault prediction and latency mitigation. Without the overhead of training bespoke machine learning models, pre-deployed LLM agents interpret structured telemetry streams—CPU load, network jitter, queue depth, and disk I/O—to generate actionable insights. These agents communicate with cloud services such as Azure Monitor, Open Telemetry, and Kubernetes to trigger automated failover and scaling actions. The architecture integrates FPGA (Field-Programmable Gate Array) acceleration and kernel bypass technologies for deterministic low-latency processing. This approach yields a scalable, explainable, and operationally efficient framework for real-time fault recovery in HFT environments.