Repetitive visual checks
Inspectors look at near-identical items all shift. Attention naturally varies, and two people may judge the same borderline mark differently.
In development · Pilot enquiries open
Make every inspection easier to review.
InspectraWorks helps quality teams organise production images, train inspection models on their own examples, and route flagged items to the people who make the call.
- Built for packaging, assembly and surface inspection
- Human reviewers stay in the loop on every flag
Illustrative inspection preview
See a flag the way a reviewer would.
Compare an approved reference image with a new capture, change the review threshold, and decide on each highlighted region. All images, regions and scores here are demonstration content — not output from a trained model.
Reference
Current capture
Regions scoring at or above the threshold are routed to the review queue. Lower thresholds send more items to people; higher thresholds send fewer. Choosing it is a trade-off your team validates on real data.
The problem
Visual inspection is careful work with scattered evidence.
Most quality teams already know what a defect looks like. The hard part is checking the same thing thousands of times, keeping the evidence, and seeing patterns across shifts and lines.
Scattered inspection records
Photos live on phones, shared drives and line PCs. Findings sit in spreadsheets or paper forms, separate from the images that justify them.
Hard-to-see defect patterns
Without consistent labels and history, it is difficult to tell whether a defect is rising, which line it comes from, or whether a fix worked.
Features
One place for images, labels, models and decisions.
The planned feature set covers the full loop from first capture to trend report. Interface previews below are design mock-ups.
Production image collection
Bring in images from line cameras or manual uploads, tagged with line, station, product and batch so every image has context.
Defect annotation
Draw boxes and assign labels from your own defect catalogue, so models learn your definitions rather than generic ones.
Model training workflows
Guided steps to split data, train, and validate against held-out images, with results presented for your team to review before use.
Flagged-image review queues
Flagged items go to a queue where reviewers confirm, reject or escalate, with the image, region and reference side by side.
Inspection history
Every image, flag and reviewer decision kept together and searchable by line, product, batch, label and date.
Defect trend reporting
See confirmed defects by type, line and period to spot rising issues and check whether corrective actions had an effect.
How it works
Four steps, with people at both ends.
-
01
Collect representative images
Capture good and defective items under real line conditions — the lighting, angles and product variants you actually run.
-
02
Label examples
Your quality experts mark defects using your own categories. Clear, consistent labels matter more than volume.
-
03
Train and validate models
Train on labelled images and test against images the model has not seen. Your team reviews the results before deciding how to use it.
-
04
Review flagged items
New captures are scored and flagged items go to reviewers. Their decisions build history and become future training examples.
InspectraWorks is designed to support inspection decisions, not make release or rejection decisions on its own.
Planned AWS architecture
Cloud by default, local where the line needs it.
The planned design uses managed AWS services. It is a working plan and may change as development and pilots progress.
Amazon S3 for images
Production images, annotations and dataset versions stored in S3, with access scoped per customer workspace.
Amazon SageMaker AI for training and inference
Training and validation jobs run in SageMaker AI. Approved models can be served from cloud inference endpoints for new captures.
Amazon ECS for application services
The web application, review queues, inspection history and reporting APIs run as containerised services on ECS.
Optional: AWS IoT Greengrass at the line
Where connectivity is limited or a result is needed quickly at the station, a trained model can be packaged and run on a local Greengrass device. Results and flagged images sync to the cloud for review when a connection is available. Whether local inference is appropriate is assessed per site.
Validation
Performance depends on your line, so we measure it there.
An inspection model only knows the conditions it was trained and tested on. We don't publish accuracy figures, because a number from one factory says little about another. Results are established during a pilot, on your images, with your reviewers.
- Lighting. Glare, shadows and changes between shifts can hide marks or create false ones.
- Cameras and optics. Resolution, focus, angle and lens choice set the smallest detail that can be seen.
- Product variation. New SKUs, colours, suppliers or materials can look different from training examples.
- Production-specific training data. Models need representative examples from your own line, including rare defects.
A pilot sets success criteria with you up front, measures results on held-out images, and keeps human review in place throughout.
About
Built around the people who already do inspection well.
InspectraWorks started from a simple observation: quality teams rarely lack expertise — they lack a consistent way to capture it. Their judgement is spread across inspectors, shifts and spreadsheets.
We're building a platform that keeps images, labels, models and reviewer decisions together, so that judgement can be applied consistently and improved over time. Reviewers stay responsible for quality decisions; the software helps them focus attention where it is needed.
- Core workflowsImage collection, annotation and review queue in development
- Training & validation pipelineBeing built on the planned AWS architecture
- Pilot programmeAccepting enquiries from manufacturers and QA teams
- General availabilityNot yet scheduled
InspectraWorks is not yet generally available and holds no product certifications.
FAQ
Questions quality teams ask us.
Is InspectraWorks available today?
Not yet. The platform is in development and we are speaking with manufacturers about pilots. Pilot scope, timing and terms are agreed individually.
Will it replace our inspectors?
No. It is designed to help reviewers focus on the items that need attention. Flagged items are routed to people, and your team keeps authority over pass, hold and reject decisions.
How accurate is it?
We don't quote accuracy figures. Performance depends on lighting, cameras, product variation and the training data from your production line. A pilot measures results on your own held-out images against criteria we agree together.
Does it guarantee every defect will be found?
No inspection method — manual or automated — finds every defect. Models can miss defects and flag acceptable items. That is why review queues, thresholds and inspection history are central to the design.
What cameras do we need?
We expect to work with images from existing line cameras where they are suitable. Part of a pilot is checking whether your current images show the defects you care about clearly enough, and recommending changes if not.
How many images do we need to start?
It varies with the product and defect types. Start with representative images of good and defective items under normal conditions. Consistent labels on a smaller set are more useful than a large, inconsistently labelled one.
Can it work without a reliable internet connection?
The planned design includes optional local inference using AWS IoT Greengrass for lines with limited connectivity or tight response-time needs. Results sync to the cloud for review when a connection is available.
Where will our images be stored?
The planned architecture stores images in Amazon S3. Region, retention and access arrangements will be defined with each pilot customer.
Do you hold any certifications?
Not at this stage. InspectraWorks is in development and does not currently hold product or compliance certifications.
Pilot enquiry
Discuss your inspection workflow.
Tell us about your line and the defects you look for. We'll use it to judge whether a pilot is a good fit — and say so if it isn't.
- A conversation about your current inspection process
- A review of sample images, if you can share them
- An honest view of fit, scope and what a pilot would measure