GPU edge Linux
NVIDIA Jetson
GPU-scoped VLM / VLA packages
GPU-appropriate models under the same outbound management and provenance contract.
Bring the edge hardware you already trust.
The reference controller exists so there is a known-good target. It is not a requirement. What stays constant across every device is the contract: a manifest-scoped runtime, a bounded field edge, and admission before install.
Right-sized execution, not one runtime everywhere.
A conveyor counting pallets does not need a vision-language model. Matching the model to the device is the point — and the fit coupon enforces it.
Reference design
Indagate QCS6490
Local multimodal, OLED status
The reproducible reference hardware. Available, never required.
ARM64 Linux
Raspberry Pi 5
OpenCV, compact CPU models
Motion, edges, contours and lightweight task models where a larger model is unnecessary.
x86_64 or ARM64
Your Linux or Yocto image
Manifest-scoped runtime
Integrate an existing gateway or product image and select models inside its measured envelope.
Deterministic field edge
Zephyr companion MCU
Signals, filters, bounded rules
Normalize PLC, sensor and field-I/O telemetry without moving safety authority out of rated controls.
AI compute and industrial control, designed together.
- Compute
- QCS6490 — local multimodal runtime
- Host OS
- Linux or Yocto, ARM64 or x86_64
- Vision
- Device-scoped VLM, compact models, OpenCV
- Field edge
- Zephyr MCU, 24 V I/O, configurable safe state
- Mounting
- Wall or DIN, passive cooling
Camera trigger and strobe, separate power domains and an MCU heartbeat are part of the design. Outputs stay disabled through boot.
Tell us what you already have.
Most sites already own most of what this needs. Send the sensors and compute on your floor with your SOPs, and the free audit tells you what fits and what does not.