01 · ON THE LINE
VIAI Edge
AI inspection that runs on your line.
The edge appliance and its Edge Inspector software find each part, score it, send the reject signal and keep every verdict with its image. No server room, no GPU rack, no cloud round-trips.
<50 ms
Per-part decision
NVIDIA
Edge GPU on site
24V GPIO
Reject signal to your line
100%
Runs offline
PROBLEM → SOLUTION
What it fixes on your line.
PROBLEM
Tired inspectors miss defects late in the shift.
SOLUTION
Live View
Every part boxed, tracked and scored the same way, every shift.
PROBLEM
Several small defects, and no clear call.
SOLUTION
Severity fusion
Every defect weighed together. Critical always rejects; uncertain calls go to a human.
PROBLEM
No proof when a customer disputes a reject.
SOLUTION
History
Every verdict kept with its image, score and defect type. Export to Excel or ZIP.
PROBLEM
Every changeover means resetting the inspection.
SOLUTION
Recipes
One recipe per product. Switching products means picking a recipe.
PROBLEM
A network outage stops cloud-based tools.
SOLUTION
Offline by design
All inspection runs on the appliance. The line keeps running.
HOW IT WORKS
A senior QA inspector's reasoning, automated.
01
SEE
Capture every part
Each part is tracked with its own ID as it crosses a locked region of interest.
02
WEIGH
Score every defect
Every defect is rated by severity, size, location and confidence, then combined into one score.
03
DECIDE
Pass, review or reject
Only a reject fires the reject output. Every verdict is saved with its image.
DECISION LOGIC
Many defects. One clear call.
PASS
Part continues.
One minor defect within tolerance still passes.
REVIEW
Image saved, sent to a human.
Two or more minor defects, or an uncertain call.
REJECT
Auto-rejected with evidence.
Any critical defect is always rejected.
DETECTED ON THIS UNIT
MINOR
surface scratch
0.46
MAJOR
low fill level
0.84
CRITICAL
cracked weld
0.91
SEVERITY FUSION · FINAL DECISION
REJECT
Fused 0.94. Critical defect, so reject regardless.
SPECIFICATIONS
Specifications, not slogans.
Per-part decision
Under 50 ms
Footprint
Fits in a control cabinet
Install
DIN rail, wall or enclosure. Set up in an afternoon.
Camera I/O
USB 3 and GigE industrial cameras, RTSP and IP cameras
Line output
24 V GPIO reject signal, PLC, MQTT, REST
Compute
NVIDIA edge GPU on site, accelerated inference
Models
PatchCore and EfficientAD anomaly models, plus a defect-type classifier
Data
Every verdict stored on the appliance with image, score and defect regions
Model delivery
Signed-off bundles from VisionLab, each file checked by SHA-256
Interfaces
Browser UI, REST API with OpenAPI docs, WebSocket feed, MJPEG stream
FEATURES
What you get.
01
Live View
One screen for the shift: every part boxed, score against thresholds, yield and items per minute.
02
History
Every verdict kept with image and score. Filter by shift, product, defect. Excel and ZIP export.
03
One recipe per product
Model, defect classes with severity and action, thresholds. Switching products means picking a recipe.
04
Severity fusion
Critical defects always reject. Uncertain calls go to a human.
05
Offline by design
A software fault or network outage never stops production.
06
Accessories
Cameras, lighting, mounts and stack lights matched to your part and line rate.
SCREENSHOTS
See it in use.
History filtered to rejects, with photo evidence
Recipe detail: defect classes, severity and thresholds
USE CASES
Where plants use it.
QA MANAGER
End-of-line final inspection
Every unit checked before it is packed and shipped.
PLANT MANAGER
In-process checks
Catch defects after welding, molding or filling, before more value is added.
OPERATOR
Shift inspection
Pick product, name and shift, start the session. The screen does the rest.