Maker Mukesh Sankhla has published an Edge AI-Based Multi-Stage Visual Inspection System on Hackster.io. It is a six-slot rotating station that photographs printed circuit boards and checks them with vision models running on the board itself. He built it from "the Arduino UNO Q, a webcam, and a servo motor." This is a documented maker build, not a product for sale. The only cost figure is in the creator's opening question: "What if a $60 SBC could turn a simple webcam into a complete industrial vision inspection system?" Arduino featured the build on its blog on 15 October 2026, in a post titled "Building a super-affordable automated PCB inspection system".
What the station does
A servo turns a platter with six slots and brings each board under an overhead webcam. Inspection happens in stages. The system first classifies the slot as empty, holding a bare PCB, or holding an assembled PCBA. Bare boards are then checked for scratches and visible damage. Assembled boards get component-level inspection, which covers large parts such as USB-C and JST connectors, switches, power ICs and inductors, and smaller 0805 LEDs, resistors and capacitors. Each component is marked as present or missing.
That work is spread across four vision models: Slot Occupancy, PCB Inspection, PCBA Big Component and PCBA Small Component. Slot Occupancy is an image classifier. The other three are object-detection models. Sankhla labelled his images with the Make Sense annotation tool. The models were "trained separately using YOLO and then brought into Edge Impulse using Bring Your Own Model (BYOM)". For the large-component model he used YOLO11 Nano, trained on a PC and exported to ONNX before the import. He reports 97.94% precision, 100% recall and 99.50% [email protected] for that model on its validation set. The example dataset behind those figures was small: 44 images, 36 for training and 8 for testing. His test subject was a PCB of his own design, a custom battery charging and discharging circuit board.
Splitting the work across the UNO Q
The build depends on the UNO Q's two-processor layout. Arduino's product page describes the board as combining a "Linux Debian-capable Qualcomm Dragonwing QRB2210 microprocessor (MPU)" with a "real-time STM32U585 microcontroller (MCU)". The QRB2210 has a quad-core Arm Cortex-A53 CPU at 2.0 GHz and an Adreno GPU. The board comes with a choice of 2GB or 4GB of RAM, and it carries 4 RGB LEDs and an 8 x 13 LED matrix.
Arduino's post explains how the station uses that split: "The STM32 on the UNO Q is able to control the servo movement, while the Linux side can run the vision models." Training happens on a PC, but inference runs on the UNO Q, so the station works without relying on the cloud.
The hardware
Besides the UNO Q, the main parts are:
- DF DSS-M15S 360° servo, which turns the platter
- Logitech C270 webcam, with its camera PCB taken out of the housing and mounted in the station
- WS2812B RGB LED for status indication
Sankhla designed the enclosure in Fusion 360 as five 3D-printed parts: Body, Rotating Plate, Pole, Pole Cover and Dome Diffuser. The pole is attached to the body with a separate Pole Mount. He printed the parts in PLA Pro on a Bambu Lab P1S, and the Hackster page links to the Fusion 360 design. Arduino notes that the whole structure is 3D-printable, that the turntable can be shaped to cradle the boards being inspected, and that the number of stations can be changed. Its example is a four-station version that would inspect both sides of unpopulated boards and both sides of populated boards.
Who it is aimed at
Arduino describes a gap between one-off prototypes and mass production, where small teams and businesses make PCBs "in relatively low quantities – say dozens or hundreds a month." It says inspection needs in that market are "under-served" and presents Sankhla's station as a very affordable automated answer.
The models are trained on one specific board. Sankhla says the same approach "can be applied to almost any PCB or PCBA depending on your inspection requirements". He also recommends capturing training images with the same camera and setup that will be used for inspection. On his board the components sit in fixed positions, so he turned off horizontal and vertical flipping during augmentation to avoid teaching the model the wrong spatial layout.
What It Means for Makers
- Who it suits: small teams making their own PCBs in the dozens-to-hundreds-per-month range who want a repeatable visual check of bare and assembled boards.
- Where the effort goes: mostly into data. Each board you want to inspect needs its own captured and labelled images and its own trained models before deployment through Edge Impulse.
- Hardware outlay: an Arduino UNO Q (2GB or 4GB RAM), an off-the-shelf webcam, a 360° servo, one RGB LED and a 3D printer for the enclosure.
- Design files and documentation: the full write-up, the Fusion 360 design link and the training workflow are on Hackster.io. Follow the creator's documentation for assembly and wiring, and for anything to do with the battery circuit used as the test board.
