Vision AI that catches defects nobody taught it to find.

Onqyra is an edge-native vision foundation model for the factory floor. Show it a dozen good parts and it flags anomalies it was never trained on — then traces every defect class back to the process that caused it.

LIVE
Simulated feed
inference
pass 0anomaly 0rate 0.0%
edge · on-prem · 0 frames uploaded
Inspection log
  • Waiting for parts…
Root-cause signal

No correlated pattern yet. Watching Line 3 for process drift.

A dozen good samples
That's all it takes to teach a new part number. No labelled defect library, no rules engineer.
Zero-shot anomalies
Flags scuffs, voids, contamination and drift it has never seen — because it learned what good looks like.
Closed-loop root cause
Correlates each defect class with line, shift, tool and batch, and pushes the signal to process control.

Built for the high-volume, high-defect-cost lines everyone else ignores.

Teach. Detect. Trace.

Three moves that replace a rules engineer, a labelled defect library and a fatigued second-shift inspector.

Teach it with a handful of good parts.

Point a camera at the line, capture a dozen known-good samples, and Onqyra builds a self-supervised model of what normal looks like for that part number. No defect library. No rules engineer. New SKU on Monday, inspecting by lunch.

  • Few-shot learning from good samples only
  • One model per part, versioned and rollback-safe
  • Re-teach in minutes when the product changes
teach · blister-pack-v14simulated
good_01.png
good_02.png
good_03.png
good_04.png
good_05.png
good_06.png
good_07.png
good_08.png
good_09.png
good_10.png
good_11.png
good_12.png
Model ready
12 samples · self-supervised · no labels required
Deploy to Line 3

Detect anomalies it was never trained on.

Because the model learned normal, anything that deviates lights up — a scuff, a void, a foreign object, a shade drift. It localises the defect on the part, scores it, and clusters new defect types automatically so your team can name them later.

  • Zero-shot anomaly detection with pixel-level localisation
  • Auto-clustered defect classes, human-named in one click
  • Runs at line speed on the edge appliance
detect · line 3 · cam 02simulated
anomaly map · pixel-levelscore 0.97
  • Missing tablet
    named by QA
    0.97
  • Foil scuff
    named by QA
    0.91
  • Unclassified cluster #7
    never seen before
    0.88
  • Seal void
    named by QA
    0.84

Trace every defect class back to the process.

Each anomaly is stamped with line, shift, tool, batch and upstream sensor state. Onqyra correlates defect classes with process events and pushes root-cause signals to your PLC, MES or SCADA — so the line fixes itself instead of just rejecting parts.

  • Correlation across line, shift, tool change, supplier lot
  • Signals over OPC-UA, MQTT, Modbus or REST
  • Scrap and warranty impact measured in weeks, not quarters
trace · foil scuff · last 72 hsimulated
Where “Foil scuff” shows up
lift vs. plant baseline
  • Line 3 · after 06:40 tool change3.4×
  • Supplier lot B-118 foil2.1×
  • Night shift · sealer temp > 182 °C1.6×
  • Line 1 · baseline1.0×
Signal pushed · OPC-UA

Foil scuff correlates with Line 3 after tool change. Recommend sealer roller inspection before next changeover.

From “defect detected” to “here’s why.”

Rejecting a bad part is table stakes. Onqyra stamps every anomaly with the state of the line and finds the process event that explains it — then closes the loop.

  • Process-aware
    Line, shift, tool change, batch, supplier lot and upstream sensor readings ride along with every frame.
  • Correlation, not just counting
    Defect classes are tested against process events continuously. When one lifts, you know within the hour.
  • Closed-loop signals
    Root-cause findings publish to PLC, MES and SCADA over OPC-UA, MQTT, Modbus or REST — so the line can react.
Foil scuff rate · Line 3
% of parts · hourly · last 36 h · synthetic data
3.4× lift after tool change
NightDayEveningNightDay0%1%2%3%4%5%12:0018:0000:0006:0012:0018:0023:00tool change · 06:4009:00 · after3.3% scuff rate
Hover to inspect. Signal published to MES at 07:35 · recommend sealer roller check before next changeover.

The second shift never gets tired. Neither should your inspection.

Human inspection fatigues. Rules-based machine vision fossilises. Onqyra learns.

Human inspectionRules-based machine vision Onqyra
Setup for a new part number
Train an inspector, hope they retain it
A rules engineer, per SKU, per change
A dozen good samples. Minutes.
Catches defect types nobody defined
Sometimes — if they're looking
Only what was explicitly programmed
Zero-shot anomaly detection
Consistency across a shift
~70–80% catch rate, fades within hours
Consistent, until the product changes
Every part, every shift, same eyes
Survives a product change
Retraining and a learning curve
Breaks. Re-engineer the rules.
Re-teach from new good samples
Explains the root cause
Anecdotal, on a good day
Pass/fail only
Correlated to line, tool, shift, lot
Cost model for a mid-sized plant
Headcount per line per shift
Capex plus integrator days
From $200 per camera per month

Runs on the line. Stays in the plant.

Inference happens on an edge appliance next to the camera, not in someone else's cloud. Your frames, your models, your network — with the connectivity a real plant floor demands.

deployment topologyplant network · VLAN 40
Cam 01
GigE · line 3
Cam 02
GigE · line 3
Cam 03
USB3 · line 4
Onqyra Edge
on-device inference
PLC / SCADA
verdict · OPC-UA
MES / historian
root cause · MQTT
0 frames leave the plant
Any industrial camera
GigE Vision, USB3 Vision, line-scan or area-scan. Keep the optics you already own.
Edge-native inference
Models run on the Onqyra edge appliance beside the line, at line speed, with no cloud round-trip.
Frames never leave the plant
Images stay on-prem by default. Only anonymised metrics and model deltas sync — or nothing at all, air-gapped.
Talks to what you run
OPC-UA, MQTT, Modbus TCP, PROFINET via gateway, and REST for MES, SCADA and historians.
Fleet-managed models
Teach on one line, roll out to twelve. Every model versioned, staged and reversible.
Built for audit
Every verdict, frame hash and root-cause signal is logged and exportable for QA and regulatory review.

From $200 per camera per month.

Priced like software, not like a capex project. Scrap and warranty reduction show up within weeks — which is why pilots convert.

Per camera

$200
per camera / month

Up to $800 depending on resolution and frame rate.

  • Few-shot teaching, unlimited part numbers
  • Zero-shot anomaly detection & localisation
  • Inspection log, exports and alerts
  • Edge appliance included on pilot
  • PLC verdict output over OPC-UA

Per line

Most pilots
Custom
per line / month

All cameras on a line, root cause and process integration.

  • Everything in Per camera
  • Root-cause correlation across line, shift, tool, lot
  • Closed-loop signals to MES, SCADA, historian
  • Fleet model management & staged rollouts
  • Named engineer during pilot

Plant & enterprise

Let's talk
multi-line, multi-site

Air-gapped deployment, validation support and SLAs.

  • Everything in Per line
  • Air-gapped and validated environments
  • Audit-grade verdict and frame-hash logging
  • Cross-site defect intelligence
  • Uptime and response SLAs

Pilot pricing and volume discounts available. Camera pricing scales with resolution and throughput; we'll quote after a 20-minute line review.

Questions from the plant floor.

The ones quality managers and process engineers ask first.

  • A handful of good parts — typically a dozen — captured on your own line with your own lighting. Onqyra learns a self-supervised representation of “normal” for that part and flags deviations. You never need a labelled defect library, and you can add named defect classes later as they appear.

Put an inspector on every line that never blinks.

Pilots start with one camera and a dozen good parts. Most plants see scrap move within the first weeks.