Introducing AI Builder: The Third Chapter Of Averroes VisionRepo
Averroes
May 13, 2026
Today we’re announcing AI Builder, the newest feature in VisionRepo and the most consequential one yet. To explain why, it helps to look at the road that got us here.
The Story So Far
Earlier this year, we launched VisionRepo, our visual data management platform for industrial AI. The vision was straightforward. Manufacturing teams were drowning in inspection images scattered across AOI systems, wafer inspection tools, photomask review stations, and a dozen other sources. None of it was organized. None of it was searchable. None of it was usable for training the AI models that could actually solve their inspection problems. VisionRepo gave teams a single place to ingest, organize, filter, version, and query massive visual datasets. Directory structures that mirror product lines and process steps. Advanced metadata filtering. Split, copy, and move operations. Dataset versioning so experiments are reproducible six months later. The foundation.
A month after launch, we introduced Auto-Labeling. This one was built specifically for large AI labs and teams sitting on massive image datasets that need labeling at scale. Here’s how it works. You start by labeling a small set of around 30 images. The platform uses that starter set to build a small model in minutes. That model then labels the rest of your dataset automatically. The results speak for themselves: better label quality than fragmented human teams can produce, consistent labeling across the entire dataset, and dramatically faster turnaround. A labeling job that used to take a team of annotators weeks now takes one engineer a single review pass. The consistency piece is what AI labs care about most. Human labelers drift. They interpret edge cases differently. They get tired and make mistakes on hour eight of a long session. A model trained on a clean starter set doesn’t drift. Once it’s labeling, every image gets the same treatment. Today, we’re introducing AI Builder. This is the chapter where everything connects. If VisionRepo gave teams the data foundation and Auto-Labeling solved the annotation bottleneck, AI Builder is where those clean, labeled datasets actually become production-grade models. Without data scientists. Without infrastructure teams. Without six-month timelines.
What AI Builder Does
AI Builder lets anyone build production-ready visual AI models themselves. Process engineers, quality engineers, metrology teams, and inspection leads on the manufacturing side. R&D groups exploring new applications. Researchers and students working on industrial vision problems. Anyone who has the data and the domain knowledge but doesn’t have a data science team behind them. Classification. Object detection. Segmentation. And anomaly detection (just released as part of this launch, more on that below).
Here’s the number that breaks the status quo: classification, detection, and segmentation models all reach upper-90s accuracy from as few as 30 labeled images per defect class. A process engineer can grab 30 images from VisionRepo, run them through Auto-Labeling, train a model in AI Builder, and have a real production-grade defect detector running before they go home.
That’s the prototyping unlock. It’s not about the model architecture. It’s about the cost of trying things. When experimentation gets cheap, manufacturing teams start experimenting. New defect class on the line? Prototype a classifier. New product introduction? Spin up a detection model and see how it holds up. Most of the value of AI in manufacturing has been locked behind the time and cost of finding out whether it works at all. AI Builder removes that gate.
Features Shipping In AI Builder
Four model types, one platform. Classification, object detection, segmentation, and anomaly detection (with a few-shot variant for rare defect classes).
Anomaly detection that needs zero defect labels. Models train exclusively on known-good images, learn what normal surface appearance looks like, and flag anything that drifts away. Exactly what photomask inspection and early-stage process qualification need, when the defect space is still being mapped.
MiniModels for edge deployment. Compact, optimized inference variants generated from the same training pipeline as the full-scale models. They run on CPU-class hardware and mobile edge devices. No GPU required at the deployment point. For distributed fabs with dozens of inspection stations, this is the difference between AI inspection at every station and AI inspection only where the infrastructure budget allowed.
Flexible GPU and CPU deployment. Heavy throughput in the data center, lightweight inference at the edge, same model lineage either way.
Video annotation and object tracking. For real-time process monitoring where temporal defect dynamics matter and a single frame doesn’t capture what’s actually happening.
REST API integration. Inference endpoints drop straight into MES, APC, and SPC systems. Defect classifications, anomaly scores, and detection outputs flow into yield analysis and process control without custom middleware.
Usage-based pricing. You pay for what you use. No oversized commitments before AI proves itself on the line.
The Three Pieces Working Together
This is the part that matters. VisionRepo, Auto-Labeling, and AI Builder were not designed as three separate products. They were designed as one workflow released in three stages.
Ingest your inspection images into VisionRepo. Let Auto-Labeling do most of the annotation work. Build, train, evaluate, and deploy your model in AI Builder. Push inference results into your existing MES, APC, and SPC systems via the API. Iterate with same-day cycles when the process changes.
That’s the workflow that used to require a data science team, an infrastructure team, an annotation team, and a deployment engineering team. Now it requires one inspection engineer and an afternoon.
Who This Is For
Process engineers who want to test whether AI can catch a defect type without commissioning a quarter of work to find out. Quality and yield engineers dealing with manual review backlogs because legacy AOI systems throw too many false rejects and miss too many real defects. Metrology teams who need to operationalize inspection models faster than their data science partners can deliver. Manufacturing automation leads building scalable inspection coverage across multiple lines and multiple sites.
It’s also for R&D groups exploring novel inspection applications, university researchers working on industrial vision problems, and students learning how production-grade visual AI actually gets built. The platform doesn’t care whether you’re shipping to a fab next quarter or running an experiment for a thesis. If you have images and a question, AI Builder gives you a way to answer it.
If any of that describes you, this launch is for you.
The technical capability of deep learning for industrial inspection has been ready for years. What was missing was an operational layer that put that capability in the hands of the engineers closest to the process. That layer is now complete.
Averroes AI develops AI Builder, part of the VisionRepo platform for industrial visual AI.
Today we’re announcing AI Builder, the newest feature in VisionRepo and the most consequential one yet. To explain why, it helps to look at the road that got us here.
The Story So Far
Earlier this year, we launched VisionRepo, our visual data management platform for industrial AI. The vision was straightforward. Manufacturing teams were drowning in inspection images scattered across AOI systems, wafer inspection tools, photomask review stations, and a dozen other sources. None of it was organized. None of it was searchable. None of it was usable for training the AI models that could actually solve their inspection problems. VisionRepo gave teams a single place to ingest, organize, filter, version, and query massive visual datasets. Directory structures that mirror product lines and process steps. Advanced metadata filtering. Split, copy, and move operations. Dataset versioning so experiments are reproducible six months later. The foundation.
A month after launch, we introduced Auto-Labeling. This one was built specifically for large AI labs and teams sitting on massive image datasets that need labeling at scale. Here’s how it works. You start by labeling a small set of around 30 images. The platform uses that starter set to build a small model in minutes. That model then labels the rest of your dataset automatically. The results speak for themselves: better label quality than fragmented human teams can produce, consistent labeling across the entire dataset, and dramatically faster turnaround. A labeling job that used to take a team of annotators weeks now takes one engineer a single review pass. The consistency piece is what AI labs care about most. Human labelers drift. They interpret edge cases differently. They get tired and make mistakes on hour eight of a long session. A model trained on a clean starter set doesn’t drift. Once it’s labeling, every image gets the same treatment.
Today, we’re introducing AI Builder. This is the chapter where everything connects. If VisionRepo gave teams the data foundation and Auto-Labeling solved the annotation bottleneck, AI Builder is where those clean, labeled datasets actually become production-grade models. Without data scientists. Without infrastructure teams. Without six-month timelines.
What AI Builder Does
AI Builder lets anyone build production-ready visual AI models themselves. Process engineers, quality engineers, metrology teams, and inspection leads on the manufacturing side. R&D groups exploring new applications. Researchers and students working on industrial vision problems. Anyone who has the data and the domain knowledge but doesn’t have a data science team behind them. Classification. Object detection. Segmentation. And anomaly detection (just released as part of this launch, more on that below).
Here’s the number that breaks the status quo: classification, detection, and segmentation models all reach upper-90s accuracy from as few as 30 labeled images per defect class. A process engineer can grab 30 images from VisionRepo, run them through Auto-Labeling, train a model in AI Builder, and have a real production-grade defect detector running before they go home.
That’s the prototyping unlock. It’s not about the model architecture. It’s about the cost of trying things. When experimentation gets cheap, manufacturing teams start experimenting. New defect class on the line? Prototype a classifier. New product introduction? Spin up a detection model and see how it holds up. Most of the value of AI in manufacturing has been locked behind the time and cost of finding out whether it works at all. AI Builder removes that gate.
Features Shipping In AI Builder
Four model types, one platform. Classification, object detection, segmentation, and anomaly detection (with a few-shot variant for rare defect classes).
Anomaly detection that needs zero defect labels. Models train exclusively on known-good images, learn what normal surface appearance looks like, and flag anything that drifts away. Exactly what photomask inspection and early-stage process qualification need, when the defect space is still being mapped.
MiniModels for edge deployment. Compact, optimized inference variants generated from the same training pipeline as the full-scale models. They run on CPU-class hardware and mobile edge devices. No GPU required at the deployment point. For distributed fabs with dozens of inspection stations, this is the difference between AI inspection at every station and AI inspection only where the infrastructure budget allowed.
Flexible GPU and CPU deployment. Heavy throughput in the data center, lightweight inference at the edge, same model lineage either way.
Video annotation and object tracking. For real-time process monitoring where temporal defect dynamics matter and a single frame doesn’t capture what’s actually happening.
REST API integration. Inference endpoints drop straight into MES, APC, and SPC systems. Defect classifications, anomaly scores, and detection outputs flow into yield analysis and process control without custom middleware.
Usage-based pricing. You pay for what you use. No oversized commitments before AI proves itself on the line.
The Three Pieces Working Together
This is the part that matters. VisionRepo, Auto-Labeling, and AI Builder were not designed as three separate products. They were designed as one workflow released in three stages.
Ingest your inspection images into VisionRepo. Let Auto-Labeling do most of the annotation work. Build, train, evaluate, and deploy your model in AI Builder. Push inference results into your existing MES, APC, and SPC systems via the API. Iterate with same-day cycles when the process changes.
That’s the workflow that used to require a data science team, an infrastructure team, an annotation team, and a deployment engineering team. Now it requires one inspection engineer and an afternoon.
Who This Is For
Process engineers who want to test whether AI can catch a defect type without commissioning a quarter of work to find out. Quality and yield engineers dealing with manual review backlogs because legacy AOI systems throw too many false rejects and miss too many real defects. Metrology teams who need to operationalize inspection models faster than their data science partners can deliver. Manufacturing automation leads building scalable inspection coverage across multiple lines and multiple sites.
It’s also for R&D groups exploring novel inspection applications, university researchers working on industrial vision problems, and students learning how production-grade visual AI actually gets built. The platform doesn’t care whether you’re shipping to a fab next quarter or running an experiment for a thesis. If you have images and a question, AI Builder gives you a way to answer it.
If any of that describes you, this launch is for you.
Get Started
Explore AI Builder to see the platform in action.
The technical capability of deep learning for industrial inspection has been ready for years. What was missing was an operational layer that put that capability in the hands of the engineers closest to the process. That layer is now complete.
Averroes AI develops AI Builder, part of the VisionRepo platform for industrial visual AI.