Inference on the device.
Models, control and cryptography on the hardware, measured against its deadline.
Atrafficcontroller,arelayoraroadsideunithasacycletime,apowerbudgetandasafetycase.Webuildwhatfitsinside:quantisedmodelsonNPU,real-timeLinux,controlthatcanrefuseanunsafeaction,cryptographymeasuredonthemicrocontroller.Everynumberistakenontheboard.
On-device inference
int8 models on NPU silicon with TFLite and Vela. Latency measured on the target, not in a simulator.
Real-time Linux and TSN
Yocto, PREEMPT_RT, TSN/gPTP, OP‑TEE. Built from source. Reproducible.
Learned control at the edge
The policy proposes. A conformance mask vetoes what the standard forbids. A watchdog takes over if the node goes quiet. Validated against a physical traffic controller. Test kit public (Apache-2.0).
Cryptography under deadlines
Post-quantum signatures measured on Cortex-M and Cortex-A against IEC 61850 GOOSE and V2X budgets. Benchmarks public. V2X results in IEEE Access.
Frequently Asked Questions
Yes, within limits we measure. int8 detection and classification runs on Ethos-U class NPUs. Memory is the ceiling, so model and tensor arena are sized together.
Digital twin first, then hardware-in-the-loop against a physical controller, then the street. At every stage a conformance mask and a watchdog can veto the policy.
Some schemes fit, some do not, and it changes with the core. We measure on the target and publish the benches. The V2X case is in IEEE Access.
Cortex-M microcontrollers, Cortex-A processors with NPUs, NVIDIA Jetson for perception, multi-band GNSS receivers. We bring up the BSP when a real-time kernel is needed.