Designing Embedded Intelligence through Co-Design Substrates Optimization and System Assurance

Embedded intelligence places sensing, inference, adaptation, and decision making inside physical systems limited by energy, memory, latency, cost, and reliability. This narrative review examines the field along four connected axes: hardware and software co-design, computational substrates, optimization at the model, mapping, and system levels, and applications in industrial automation, automotive electronics, healthcare monitoring, and the Internet of Things. Timing, safety, and security are treated as requirements that cut across all four axes. The review separates embedded intelligence from constrained on-device machine learning and from edge artificial intelligence, and it frames deployment as a cross-layer problem in which model quality, hardware mapping, timing, energy, security, and adaptation constrain one another. Two literature-based case studies, one on a field-programmable gate arrayand one on distributed microcontrollers, show how a gain at one layer produces a cost at another. A recomputation of the efficiency ratios reported in one of the studies shows why measurement boundaries and metric definitions must be declared. Six research priorities are derived from limitations that recur in the reviewed sources: lifecycle energy accounting, assurance of learned components, security of distributed deployments, reproducible benchmarking, portable deployment toolchains, and concept drift after deployment. Each priority is paired with a research direction and evaluation criteria for later studies. The evidence consists mainly of surveys, analytical studies, and laboratory implementations, so quantitative values are reported as stated in the sources, recomputed values are labeled as such, and approaches are not ranked. The proposed formulations and criteria are conceptual.