AI-on-the-edge refers to deploying machine-learning models directly on resource-constrained devices that operate close to the data source, enabling low-latency inference and enhanced privacy. As these devices often run tiny ML workloads on microcontrollers or specialized SoCs, efficient use of memory, compute resources, and energy becomes essential. Edge AI systems must therefore balance model accuracy with strict power budgets, limited storage, and real-time constraints. Another fundamental challenge is security: executing neural networks on devices physically accessible to end users increases the risk of model theft, tampering, and data leakage. Hardware accelerators and trusted-execution components can help enforce confidentiality by providing secure cryptographic operations and isolated execution paths.