Dataset
Client dataset annotation.
IA-06 · Applied AI
Computer vision
Useful when there is volume and consistency: invoices, labels, defects, counting, occupancy. YOLO-v9 fine-tuned + GPT-4V for open-ended cases.
Computer vision has a clear sweet spot: high-volume tasks with consistent input images and clear economic impact — production defect inspection, OCR on unstructured invoices, warehouse unit counting, occupancy monitoring.
We work with YOLO-v9 fine-tuned on the client's dataset for closed cases (specific defects, defined counts) and with GPT-4V when the input is more open-ended (ambiguous visual classification, text extraction from variable forms).
In the most recent industrial deployment — weld defect detection — we achieved 99.2% recall on critical defects with 2.1% false positives, and a 14% increase in line throughput. Human inspectors moved to qualitative review and training.
Defect · weld · 99.3%
Weld OK · 97.1%
The Computer vision process.
The process · four stagesTraining
YOLO fine-tuning on the dataset.
Deployment
Edge for latency; cloud when justified.
Operation
Conservative policy. Inspectors moved to qualitative review.
What is included
What Computer vision includes.
The operational detail: what we deliver as part of the engagement and what we keep active afterwards.
-
Defect inspection
Recall >99% in closed projects.
-
Document OCR
Invoices, labels, forms.
-
Automated counting
Warehouse, occupancy, production.
-
Edge pipeline when needed
Low latency on the shop floor.
-
Triple human validation
During calibration.
-
Human redeployment
Inspectors reassigned to qualitative tasks.
Regulatory framework
The regulatory framework.
If computer vision is used for decisions about individuals, the AI Act classifies it as high risk.
EU
AI Act
Applicable to this service
EU
RGPD
Applicable to this service
ISO
42001
Applicable to this service
CE
product
Applicable to this service
Summum cluster
How it intersects with related services.
Industrial vision is AI + systems + compliance sometimes.
Frequently asked questions about Computer vision.
Do you need to annotate the dataset?
We coordinate. Professional when justified; tools such as Labelbox.
Edge or cloud?
Edge when latency matters.
What about the camera?
We do not sell hardware. We advise and integrate.