Edge Command Router & Decision Workbench

Turn machine operation requests into ranked commands, all computed at the edge.

A local web workbench that turns short machine requests into a ranked command choice on RK3576 and RK3588.

SenseCraftOn-Device LMAI Agent Ready

Reusable Site Templates

Conveyor lines, automated workstations, and industrial control panels requiring local command resolution.

Conveyor line dispatch

Conveyor line dispatch

Ranks allowed commands for natural-language speed and stop requests within a predefined safe domain.

Production inspection queries

Production inspection queries

Operators input status check phrases, and the local model immediately identifies the corresponding action.

Industrial gateway action validation

Industrial gateway action validation

Supervisory systems send candidate actions; the edge host evaluates and ranks them to prevent out-of-bounds execution.

Requires RK3576 or RK3588 host with Docker and librknnrt.so runtime installed.

How it works

The edge IPC receives control requests, the RK3588/RK3576 NPU scores candidate options via on-device models, and outputs structured choices via HTTP POST.

The system deploys as a lightweight single container with a web-based operator workbench and an RKLLM-backed decision engine. Requests evaluate strictly against candidate actions in a whitelist, returning ranked JSON payloads via RESTful API (POST /api/decide).

Edge Computing Platform

Runs on RK3576 / RK3588 SoCs (such as reComputer RK3588 series and LubanCat3), leveraging dedicated NPU hardware without discrete graphics cards.

How well it works

Capabilities available to operators in the web UI and to systems via REST API.

Real-time command ranking

  • Match natural language requests to actions
  • Measured p50 latency of 335.43ms
  • Returns ranked choices with confidence

Whitelist boundary safety

  • Ranks only caller-provided candidate actions
  • Prevents out-of-domain command hallucination
  • Supports customizable prompts and options

On-device offline execution

  • Models run completely on the local NPU
  • Factory requests never leave the internal LAN
  • Shipped as a self-contained Docker container

How to deploy

Check the setup and its product list, then follow the deployment guide.

Hardware Sizing & Deployment Setup

Select site conditions to generate recommended bundles and guides, or let an AI assistant estimate hardware and deploy for you.

Self-service Consult & Deploy
Deployment Options

RK3576 Kev RKLLM (2048-token deployment)

Deploy the tested RK3576 Kev RKLLM configuration with a 2048-token request guard; the engine export capacity remains 8192.

2 steps

Deployment Guide →
Target device
RK3576 with RKLLM 1.3.1

Integration Interfaces

http

Health status for the local command workbench.

/api/health · Port: 18768 · Method: GET
http

Rank candidate machine commands for a user request.

/api/decide · Port: 18768 · Method: POST
{"prompt":"Choose the best machine command.","input_text":"Read current conveyor status.","options":[{"id":"status","label":"Read status"},{"id":"stop","label":"Stop"}]}

Usage Requirements

network

Network access is needed for the first download unless the target already has the pinned artifacts.

storage

Keep at least 3GB free for a fresh Kev artifact/bootstrap operation; measured fresh peak is about 1.94GB with a 1GB reserve. The existing RKNN path documents its 4GB requirement separately.

Contact Us
We Are Glad to Be Your Hardware Partner !
Next