Intel Neural Compute Stick + ASI Biont: Edge AI and On-Device ML Without a Server

Why attach an AI agent to an Intel Neural Compute Stick?

An Intel Neural Compute Stick (NCS2) is a USB device with a Movidius Myriad X VPU that runs neural networks in parallel with a host CPU. It is one of the most accessible ways to get on-device ML: no GPU, no cloud, just a USB connector and a model converted by the OpenVINO toolkit. Intel has ended production of the NCS line, so you won't see it in new flagship products, but thousands of sticks still work in labs, factories, and university projects.

A stick alone is only a calculator. It returns tensors; it does not decide whether a defect is critical or how to react. ASI Biont adds the automation layer. When connected, Biont can request an inference, interpret the result and send commands to MQTT brokers, PLCs, Telegram bots, or any other system you already use.

Official references: Intel NCS2 product brief and OpenVINO documentation. The OpenVINO Intermediate Representation format is the official way to prepare a model for the VPU.

How ASI Biont connects to the stick

The NCS2 has no Ethernet connector, no Modbus register, and no MQTT topic. It connects over USB to an edge host. That's why ASI Biont never talks to the stick directly; it talks to the computer that owns the stick. Choose the transport based on where that computer is:

Method Typical case Python library used by ASI Biont
HTTP API Edge host and Biont are in the same LAN aiohttp
MQTT Edge host is behind NAT or on a 4G network paho-mqtt
WebSocket Live video frames must be processed in a stream aiohttp WebSocket
SSH Need to deploy a script on the host and run it remotely paramiko
execute_python Custom glue logic for an unusual host setup generated Python

For a fixed installation on the same local network, an HTTP API is the most predictable choice: one POST request, one JSON response. If the same NCS host sends data from a warehouse with a 4G modem, MQTT is better than trying to open incoming ports. The table above is a decision rule; you do not need to change the device, only the transport.

For classic fieldbus devices (Modbus, BACnet, Siemens S7), ASI Biont uses bridge.py from the dashboard, launched with --token=XXX --ports=COM3 --baud=115200 --rate=10, and then commands are sent via industrial_command(protocol=..., command=...). The NCS is not a fieldbus device, so the network path above is the natural integration route.

Universal connectivity via execute_python

ASI Biont does not wait for developers to add a “device plug-in”. The chat is the control panel. You describe the device and its connection parameters (IP address, port, baud rate, API key), and the
system generates a Python adapter that matches that description. The adapter is not a black box and it is not a compiled binary; it is plain Python that uses the standard library plus the specific client library from the transport table (aiohttp, paho-mqtt, paramiko, or just socket). You can open the generated file, adjust timeouts, add retries, or swap the low-level transport without restarting the platform.

For example, a conversation like this is enough:

```text
User: Add a device called

← All posts

Comments