SAP · Business AI · 0 to 1
Generative Defect Simulation
SAP Visual Inspection catches product defects on the production line automatically. I designed the generative-AI workflow that trains it, creating synthetic defects from a single clean photo instead of manufacturing faulty parts. Concept to shipped capability.
ROLE
UX Designer · SAP Business AI
YEAR
2024
COMPANY
SAP
Scope
0-to-1 workflow, concept to ship
70–90%
Less dataset preparation time*
80–95%
Lower physical defect-creation cost*
+10–25%
Model accuracy, broader coverage*
0 → 1
Concept to a shipped SAP capability
At a glance
UX Designer · SAP Business AI
Role
2024
Timeline
Intelligent-systems expert, VI product owner
Team
0-to-1 workflow, concept to ship
Platform
What I did
Designed the end-to-end generative-AI workflow for synthetic defect images.
Made synthetic data trustworthy for experts: compare, refine, control, approve.
Connected it to the ML pipeline and took the concept through leadership to ship.
Shipped in SAP's manufacturing portfolio · 70–90% less prep time
overview
Turning weeks of work into a few clicks.
SAP Visual Inspection is AI-powered quality control. Instead of people manually checking products for flaws, cameras and machine-learning models inspect each item on the production line and flag defects automatically. The process I innovated sits one step upstream: how those inspection models get trained in the first place.
And that is where the pain was. This is a case study about making a slow, expensive, expert-only process simple, without losing the experts' trust. A visual-inspection model needs thousands of defect images to learn from. Traditionally, manufacturers had to physically make defective parts, photograph them, and annotate each one by hand: slow, costly, and impossible to scale across product variants.
In 2024, as part of SAP Business AI, I designed a generative-AI workflow that produces realistic synthetic defects from a single clean product photo. I owned the end-to-end experience, from understanding manufacturing and ML workflows to designing the studio, building in the quality controls domain experts needed to trust it, and connecting it to the training pipeline. What began as an exploration moved onto SAP's roadmap and shipped as a real capability.
context
Innovation at the core of manufacturing.
SAP Business AI was the central team shaping AI experience guidelines and enabling consistent adoption across SAP's product landscape. My role was to bring UX leadership and consultation to multiple lines of business, making sure AI concepts were both strategically aligned and operationally feasible.
Visual inspection was a strong candidate for innovation. Machine-learning adoption in manufacturing is slowed primarily by one thing: dataset creation. Manufacturers spend time and money generating physical defects, rare defects are hard to capture and leave accuracy gaps, and every new product variant means starting training from zero. Synthetic data could cut the time and cost while improving accuracy. The opportunity was to unlock value right at the core of digital manufacturing.
The problem
The data was the bottleneck.
Models need thousands of high-quality images of both defective and good product states. Getting them was painful on two fronts.
Creating defective samples meant stopping or adjusting production.
Gathering rare defect types could take weeks.
Manual annotation needed skilled people and significant time.
Every new product variant meant repeating the whole process.
Inconsistent real-world image quality caused ML setbacks, and scaling across product lines meant recurring investment.
The brief was to design a radically more efficient process: lower cost, higher quality, and reusable across variants, without asking domain experts to trust a black box.
my role
From a workflow to a shipped product.
As the UX designer representing SAP Business AI, I owned the experience end to end: understanding manufacturing workflows, defect taxonomies, and ML training; partnering with the intelligent-systems specialist and the Visual Inspection product owner; designing the generative workflow itself; building in the clarity, transparency, and quality controls experts would demand; connecting the output to the ML training pipeline; aligning it with SAP's AI design guidelines; and preparing and presenting the concept for leadership.
Crucially, I designed it to ship, not just to demo. The concept moved from exploration onto SAP's roadmap and became a real capability in the manufacturing portfolio.
The core of the work
Designing a hard ML problem into a simple studio.
The complexity sat in the domain, manufacturing, defect taxonomies, and an ML training pipeline, and in the bar for trust: experts will not train a model on data they do not believe in. The job was to make all of that feel like a simple, controllable studio.
The complexity I was designing for
Thousands of images. Models need large, balanced datasets of defective and good states.
Physical, slow, costly. Making defects meant stopping production and weeks of effort.
Rare defects. The hardest cases to capture were the ones models most needed.
Repetition per variant. Every new product restarted the whole process.
Expert scepticism. Synthetic data is worthless if domain experts will not trust it.
It had to fit the pipeline. Output had to flow straight into real ML training.
How I made it a studio
One photo in. Start from a single clean image of any product component.
Generate on demand. Realistic synthetic defects, including the rare ones, in minutes.
Tune to reality. Control severity, scale, location, and frequency to match the real line.
Compare and approve. See original beside generated; approve or regenerate, so experts stay in control.
Export to training. Push a balanced dataset straight into the ML pipeline.
Works anywhere. Any product category, because users bring their own images.
I turned weeks of physical defect-making into a few clicks, without asking experts to trust a black box.
Design move 01
A studio, not a model.
The trust question: can a domain expert build a usable dataset without ML expertise?
The generative model is the engine, but the experience is a studio: upload a clean photo, choose a defect category, generate variations, and tune them. I designed the whole flow around the manufacturing expert, not the data scientist, so the people who understand the defects could create the data themselves, in language and controls that made sense to them.
Design move 02
Trust through comparison and control.
The trust question: can an expert believe a generated defect is realistic enough to train on?
Synthetic data only has value if people trust it. So I made trust a designed-in feature: every generated image sits beside the original for comparison, experts adjust severity, scale, location, and frequency to match their real production, and nothing enters a dataset without an explicit approve or regenerate. The expert stays the judge of quality, and the system earns confidence one decision at a time.
Original beside generated, with approve or regenerate on every image. Trust is built into the workflow, not assumed.
Design move 03
Straight into the pipeline.
The trust question: does this actually fit how models get trained, or is it a dead end?
A clever generator that produces orphan images would have solved nothing. I designed the output to flow directly into Visual Inspection's ML training pipeline: a balanced, labelled dataset, exported in one step, ready to train immediately. The value was not the images, it was removing the entire gap between needing data and having a trained model.
Impact
A shipped capability, and a step change in economics.
The concept progressed far beyond exploration. It earned strong leadership support, evolved through validation, and now ships inside SAP's manufacturing portfolio, where customers use it to accelerate defect-detection training with far less effort. It strengthened SAP's manufacturing offering, reinforced Business AI as a practical innovation leader, and became a reusable capability that adapts to any product category, since users bring their own images.
Reflection
This project reimagined how manufacturers prepare ML data. The most rewarding part was that it did not stop at a compelling demo: it was designed to be real, and it shipped. It is a reminder that the strongest AI ideas are the ones built to fit a real workflow and to earn the trust of the people who depend on them.
What it taught me
Principles for designing with generative AI.
This 0-to-1 work shaped how I bring generative AI into real products.
Generative AI is a workflow, not a button.
The model is the easy part. The product is the upload, the tuning, the review, and the handoff around it.
Earn an expert's trust with comparison and control.
People believe AI output they can inspect, adjust, and approve. Make judgement part of the flow.
Make the rare thing easy and the costly thing cheap.
The biggest wins target the steps that were previously slow, expensive, or impossible.
Design 0-to-1 to ship, not to demo.
A concept that fits a real pipeline and a real roadmap is worth more than a beautiful dead end.
next