Who it is for
For electronics manufacturers working in small batches with many part numbers: 20 to 500 boards, frequent changeovers, inspection often still by eye. Traditional AOI costs too much here and needs hours of programming for every new board.
Off-the-shelf hardware
An open-source CoreXY 3D printer already has almost everything we need. In Voron 2.4-style machines the bed stays still and only the gantry moves: we replace the nozzle with an optical head.
| Element |
Choice |
| Mechanics |
Open-source CoreXY frame, Klipper firmware, 3D-printed board fixture with datum pins |
| Sensor |
Raspberry Pi Global Shutter Camera: a global shutter does not distort the image if the head vibrates |
| Lens |
0.25× C-mount telecentric lens: about 14 µm per pixel, no parallax on tall parts |
| Light |
Programmable LED ring driven by a Raspberry Pi Pico, lighting recipes stored per session |
| Overview |
An iPhone on a fixed mount, to identify the board and check its orientation |
Precision does not depend on the belts: every image is realigned to the golden board before looking for defects.
Three architectures, one software
|
How it works |
When to choose it |
| L1 |
Fixed optics, the board moves on an X/Y stage |
The simplest, for the first bench |
| L2 |
Board fixed, the optical head moves on X/Y/Z |
Larger boards, no board movement |
| L3 |
Two lines in one machine that advance independently: while the head inspects one board, the other flips 180° on the platform |
Both sides inspected, maximum throughput |
One model sees, another decides
The visual model V1 aligns the board, checks presence, position and polarity of every component and measures how far each region deviates from good boards. It is deterministic and written in C++.
The decision model M1 does not look at pixels. It takes the structured results and answers typed questions with calibrated probabilities: is this a real defect? pass, rework, scrap or review? how severe? We use open-weight System One models, run on servers chosen by the customer.
Below the confidence threshold the operator decides. A rule engine stays alongside the model: every new version must beat it on the same inspections.
The M1 model and the OA service
On the bench a Raspberry Pi 5 does only what needs low latency: field plan, motion, light, trigger, image upload.
The heavy processing runs on GPU servers. That is where M1 works: the large model that receives images from every machine, improves with the data collected on the line and updates V1, the lighter model running next to the optical heads.
OA is the monitoring and alerting service: it shows the live status of every line, the board under inspection and the defects found, and alerts the operator when a decision is needed.
SaaS, on premise or hybrid
Lumaspect Inspect™ can be deployed in three ways, depending on what the customer needs.
|
Where the models run |
When it makes sense |
| SaaS |
Managed by us, on a cloud in Switzerland or Europe chosen together |
To start without your own infrastructure |
| On premise |
On servers inside the customer's plant |
When data must not leave the company |
| Hybrid |
Acquisition and fast decisions in the plant, M1 and OA in the cloud |
For low latency on the line and computing power where it is needed |
In every case, data and the models trained on your parts remain yours.
What we measure
- Missed defects, per category
- False alarms per board and operator time
- Time to set up a new board
- Total inspection time
A generic "99% accuracy" says little: 0.1% false alarms over 1,000 checks per board means one false alarm on every board on average.
Declared limits
Solder height and volume need calibrated 3D acquisition: they are not part of the first version. Before going further we run a freedom-to-operate analysis against existing patents.