Non-Gradient Neural Dynamics for
Real-Time Edge Intelligence

In one paragraph

Edge AI today either trains offline and ships a frozen model, or runs surrogate-gradient SNNs that still depend on backpropagation through time. We explore an orthogonal regime: a software inference engine designed for real-time operation on constrained hardware — no end-to-end backpropagation, on-device adaptation by construction. In the L-tier configuration (not the full configuration), the engine runs at a mean engine-step latency of 0.90 ms on an STM32F767 (Cortex-M7, 216 MHz). The architecture is spike-mappable. Algorithmic equivalence and state preservation for the spiking path have been validated in software. Hardware deployment on Pulsar and Loihi 2 is on the roadmap. The currently shipping implementation is the equivalent fixed-point implementation. It has been evaluated on standard PdM / SHM benchmarks using within-operating-condition protocols; we position the engine on form factor and online operation, not on out-detecting classical scorers. Cross-device generalisation remains an active research question.

Note: figures reflect the STM32F767 MCU deployment. An earlier abstract's 50 KB figure referred to frozen model weights only — it does not represent the full edge build, the deployment package, or shipping firmware size.