Core AI · Module 03 / Vision
Perception at
industrial speed.
Detection, classification, quality inspection and visual search — engineered for factory floors, warehouses, retail spaces and clinical imaging.
32ms
end-to-end latency
99.4%
detection accuracy
48
cameras / node
clickripple · vision · line-14 · shift 2
LIVEus-east
worker · PPE ok · 98%
asset · conveyor-A2 · 94%
defect · dent · 86%
live · 30 fps · lat 32ms
Detections · last 60s
- worker12
- asset34
- defect2
- unsafe zone0
Quality score
98.7%
+0.4% vs shift avg
What it does
Cameras are already installed. Computer vision makes them think.
cr-vision-2 turns any camera feed — factory, retail, warehouse or clinical — into a live stream of typed events. Objects, people, defects and states are named, scored and dispatched to the systems that already run your operation.
01
Real-time perception
Sub-40ms detection at 30 fps, running on-prem or at the edge with GPU or NPU acceleration.
02
Fine-tuned per line
Every deployment adapts to your line, product family and imaging conditions. No generic checkpoints.
03
Actionable events
Detections dispatch typed events into ERP, MES and SCADA — not another dashboard nobody watches.
How it works
One continuous flow from signal to outcome.
Every stage is instrumented, observable and safe to swap. No timelines. No boxes to connect by hand.
- Stage 1Capturecamera / sensorRGB, thermal, depth or hyperspectral input arrives at 30–120 fps.
- Stage 2Preprocessedge runtimeFrames are stabilised, colour-corrected and cropped to regions of interest.
- Stage 3Detectcr-vision-2Objects and events are localised with bounding boxes, masks and confidence.
- Stage 4Classifyfine-tuned headLine-specific classification for defect type, product SKU or safety state.
- Stage 5DecidecommitActions are dispatched: hold, alert operator, log, or continue — with signed evidence.
Live inspection
Every frame,
reasoned in 32ms.
vision · inspection · unit_00482
LIVEus-east
Classes
- Surface scratch92%
- Dent74%
- Colour drift51%
- Alignment33%
- Contamination12%
unit_00482
defect · dent · 0.86
action: hold + reinspect
thermaledgedepthrgb
Pipeline
- capture8ms
- preproc12ms
- detect18ms
- classify22ms
- decide27ms
Model
cr-vision-2 · 22B
fine-tuned · line-14 · v9
fine-tuned · line-14 · v9
Enterprise applications
Where computer vision earns its keep.
Manufacturing
Case · 01 / 04
Detect surface defects before parts leave the line.
AI capability
Edge vision · fine-tuned classifier · MES hook.
Business result
−82% escape defects · +14% first-pass yield.
Performance
Numbers that earn deployment.
99.4%
Detection accuracy
on production imagery
32ms
End-to-end latency
camera → decision → commit
82%
Fewer escapes
measured across 6 factories
48
Feeds per GPU
H100 · cr-vision-2 · fp8
Compatibility
Meets your stack where it lives.
Bring your own models, cloud and data. Everything speaks the same runtime contract.
Vision models
cr-vision-2SAM 2YOLOv10DINO v3Custom heads
Cameras & sensors
BaslerCognexAxisFLIR (thermal)Intel RealSenseiPhone MFi
Edge runtime
NVIDIA JetsonH100 / L40SIntel NPUApple Neural EngineCoral TPU
Systems of record
SAP ERPIgnition SCADAManhattan WMSPACSCustom MES
Frequently asked
Straight answers.
In most cases yes — we support standard RTSP, ONVIF and MFi streams, and can add edge boxes for cameras without on-board compute.
We ship an active-learning loop: uncertain frames are surfaced for review, and the model retrains weekly with drift detection built in.
Yes. Everything can run inside your plant on Jetson or on-prem GPU. Only aggregate telemetry leaves the site — and only if you allow it.
We support face and licence-plate anonymisation at capture, with policy enforcement in cr-guard-1 before any frame is stored.
A first inspection use case ships in 4–8 weeks, including line-specific fine-tuning and MES integration.
Computer Vision
Deploy this module in your enterprise.
Talk to an AI engineer. Ship a working proof-of-value in weeks, not quarters.