What exactly is the inspection gap at new product introduction?

It is the window between the start of a new product build and the moment a stable automated inspection recipe exists. During that window inspection falls back on manual checks and tribal knowledge, so escapes, rework and scrap concentrate exactly where the process is least understood and change is most frequent.

Why does traditional rule-based machine vision struggle at NPI?

Rule-based vision needs a fixed, well-characterised part and a specialist to hand-tune thresholds for each feature. At NPI the part is still changing, so every revision restarts that tuning cycle. The recipe is usually stable only once the ramp is over, which is precisely when it was needed most.

How many images are needed before inspection can start?

Far fewer than a from-scratch deep-learning project implies. Start with the OK images the line already produces, add the handful of defect examples available, and generate the rest synthetically. A first governed model can go live on a small set and then improve as real production images accumulate.

Does closing the gap require a data science team?

No. The people who should own inspection at NPI are the process and quality engineers who already know what a defect looks like. A no-code workflow lets them label, train, validate and release models themselves, with governance and version control handling the parts a data scientist would otherwise supervise.

Why run this on-premise instead of in the cloud?

New product images are among the most sensitive data a plant holds, and inspection decisions have to keep pace with line takt time. Running on-premise keeps that data inside the factory network and keeps inference latency deterministic, with no dependency on an external link during a ramp.

How does this scale to other lines and sites?

Because models, defect definitions and thresholds are versioned artefacts rather than one-off configurations, a recipe proven on the pilot line can be promoted to sister lines and other plants. Every site then inspects against the same definition of OK and NOK instead of a locally tuned interpretation.

How to Eliminate the Inspection Gap at NPI

How do you train an inspection system before defects even exist? In manufacturing, the biggest bottleneck isn't AI — it's data. This guide shows how to simulate defects before production starts and eliminate the inspection gap at NPI.

No real defects required to start
Hours, not weeks
Fully on-prem — no IP exposure

The Problem and the Solution

Problem

No defected samples exist yet

During New Product Introduction (NPI), inspection systems fail for a simple reason: there are no defected samples available yet.

Model training is delayed
QA has to rely on manual inspection

solution

Generate the defects before they occur

ReliVision addresses this by generating synthetic defects on-prem, before they occur in reality. Scratches, corrosion and voids can be simulated with controlled parameters, enabling inspection systems to be ready from day one.

A real sheet-metal part beside the same part with synthetic scratches applied. — A real part and the same part with controlled synthetic defects — the training data that would otherwise not exist yet.

Benefits

Simulating defects before production fundamentally changes how inspection systems are designed and deployed. Instead of waiting for defects to occur:

Models can be developed with minimal reliance on real NOK samples.
Inspection pipelines can be prepared and validated ahead of NPI.
Dependence on rare or hard-to-capture defect data is reduced.
Inspection shifts from reactive detection to proactive readiness, enabling faster ramp-up and more stable production launches.

Zero-Shot / Few-Shot Deployment

Models can be developed with minimal reliance on real NOK samples.

Faster Time-To-Production

Inspection pipelines can be prepared and validated ahead of NPI.

Reduced Dependency On Real NOK Samples

Dependence on rare or hard-to-capture defect data is reduced.

Workflow: How It Works

The workflow is fully integrated into ReliUI and can be completed within hours.

Import and Annotate

Import and annotate OK images and any available NOK images.

Define New Defect Types with the GenAI Module

Select regions of interest
Specify defect characteristics
Generate synthetic NOK samples

Review and Convert

Review and convert the generated images into training-ready datasets.

The result is a scalable dataset that reflects production conditions — without requiring real defect collection.

The three-stage flow in ReliUI: annotation, the GenAI defect definition module and the generated dataset. — All three stages run inside ReliUI — there is no export to an external tool and back.

Why This Matters

With ReliVision, synthetic data generation is available within the platform for on-premise use. This enables:

Native GenAI for synthetic defect generation within the platform

Controlled defect definition and variation

Fully on-prem execution with no IP exposure

With ReliVision

Inspection systems can now be designed, trained and adapted entirely within one platform — without external dependencies.

Close the gap before the ramp

By simulating defects before they appear in production, you remove one of the biggest constraints in inspection system deployment. Small datasets expand into production-ready training data, models are validated earlier, and production ramps without waiting for defects to occur.

Download ReliVision

Knowledge Hub