Technology
Proprietary AI, built for weld geometry.
Inspection AI is usually limited by data. SynSpect runs on proprietary AI models built in-house to need very little of it, which is why a new deployment takes about 10 training samples instead of thousands.
- 01Proprietary in-house models
- 02~90% fewer real samples
- 03~10 training samples to deploy
Most inspection AI waits on thousands of hand-labelled welds. SynSpect needs about 10.
The data problem
Labelled weld data is scarce.
A model that judges welds has to see many of them, including the defects. But defective welds are what a good shop works to avoid, and each example needs a qualified person to label it.
Collecting thousands of labelled welds for every joint, material, and scanner is slow and expensive. Most projects stall here.
The approach
Models that need far fewer real welds.
SynSpect's models are proprietary and built in-house around weld geometry. Where real labelled welds are scarce, data we generate in-house fills the gap, so the models understand welds before they see a customer's part.
In development, this cut the number of real samples needed by about 90%.
From install to live
Four steps to a deployed model.
- 1
Install SynSpect
On a standard industrial PC, connected to your scanner or profiler.
- 2
Train on about 10 samples
From your line, on your scanner, to fit the models to your weld.
- 3
Check
We run SynSpect against your own gauges, so you can compare the results.
- 4
Live
You decide when it goes live on the line.
- training samples to deploy
- ~10
- accuracy in our tests so far
- 99.7%
From SFI development and testing to date. Sample counts vary with joint type and scanner.
Start with one part on your line.
We run SynSpect against your own gauges, and you decide when it goes live.