Towards AI-Augmented Co-Compilation for Smart IoT Systems Through Domain-Specific LLMs

The explosion of smart, interconnected IoT devices is driving demand for higher performance, real-time processing, and self-* capabilities (self-adaptation, self-reconfiguration, self-healing) across heterogeneous platforms. CPUs, GPUs, and FPGAs now provide exceptional compute power but greatly complicate software toolchains. We envision AI-COMPILER, an LLM-driven, domain-specific framework that automates co-compilation and optimization of heterogeneous applications in smart cities, smart transportation, and real-time analytics. Tackling multi-objective goals—power efficiency, reliability, and dynamic reconfiguration—AI-COMPILER aims to transform how IoT and edge infrastructures are developed, deployed, and maintained. We examine critical issues such as generalization to new hardware, human-AI collaborative tuning, and semantic coherence across multiple DSLs, arguing that AI-COMPILER can usher in more agile, resilient IoT ecosystems through flexible, context-aware code generation and adaptation.