Agents optimize models for production environments, operate and retrain them, and design new ones.
Your role is to set the bar and review the results.
Optimize with customer data
Retrain when performance drops
Operational data for the next training cycle
Design a new model
When people think of AI agents, they usually think of agents that use finished AI to get work done.
Superb AI's agents work upstream: they design, optimize, and train the AI models that will be deployed in production.
Design · Train 
Production-Ready AI
Deploy to ProductionSuperb AI agents that build AI optimized for production deployment
Detect
DecideConventional agents that use finished AI to perform tasks
The three tracks of the Agentic Model Factory, in order of what you can apply today.
Optimizes ZERO, our industry-specific Vision Foundation Model (VFM), for your production environment.
Give the agent your data, and it adjusts ZERO's settings until accuracy improves.
Accuracy 0.610 → 0.968 +59%
499 images · 6 hr 22 min
Agents take on building and improving models such as detection models and VLMs. When an operator sets a goal, the agent directly uses Superb Platform's data selection, labeling, model training, evaluation, and deployment capabilities to choose and execute the necessary tasks.
Performance improved in under 8 minutes with 318 images
The agent handles everything from data selection to deployment and real-world validation
Agents take over model design from researchers. Drawing on published papers, the agent designs a detector's architecture, trains it, and validates it against set criteria — with no human in the loop. The resulting model beat the published state of the art in its size class.
Higher accuracy than YOLO across all model sizes
+1.7 to +3.4 AP vs. YOLO26 S/L/X at equal or lower parameter counts · achieved through 3 weeks of autonomous model evolution
All three tracks follow the same sequence. People set the criteria; the agent iterates, trying different changes and keeping only those that improve the score.
What differs is what the agent changes from round to round: ZERO's settings, Superb Platform tasks, or model architecture.
You provide your data and the criteria for what counts as better.
The agent reads what the last round got wrong and picks the next thing to try: refine the examples and instructions, change how images are read, or train further.
Target examples · What to filter out · Image tiling · Re-scoring · Retraining
Each change is re-scored on test data the model never trained on. Only gains too large to be chance are kept, and the reason for every keep and discard is logged.
+10% to +132%accuracy
14 runs on a single A100 GPU · 88.5 hours total
The operator states a goal, such as "build a detector from this data and deploy it."
The agent directly uses Superb Platform's capabilities, replacing manual operation. It selects data to identify missing scenarios, labels it, trains and evaluates the model, and deploys it. It then reviews the results and decides the next task on its own.
Data curation · Labeling · Model Training · Evaluation · Deployment
It runs test images through the deployed model, checks the results, and reports with evidence. The operator decides how far automation goes.
8 minto improve performance
The agent automates data selection, labeling, training, evaluation, and deployment
Specify the target inference speed, memory limits, and unsupported operations before design begins.
The agent designs candidates from the open literature, trains them, and validates against the set criteria.
Model architecture · Training method
Only models that clear the criteria are kept. Failed experiments are logged with their cause.
+25%accuracy
32 research loops · higher accuracy than YOLO26 at comparable model sizes
Of the 232 changes proposed by the agent, only 34 were accepted after clear improvements were confirmed on test data.
Those gains also held on test data that had not been used for training.
| Dataset | Images | Rounds | Accuracy before optimization | Accuracy after optimization | Improvement |
|---|---|---|---|---|---|
| cell-towers | 705 | 11 | 0.252 | 0.585 | +132% |
| cable-damage | 919 | 10 | 0.328 | 0.635 | +94% |
| water-meter | 499 | 9 | 0.610 | 0.968 | +59% |
| soccer-players | 114 | 8 | 0.623 | 0.942 | +51% |
| aquarium | 448 | 11 | 0.563 | 0.694 | +23% |
| soda-bottles | 1,547 | 9 | 0.806 | 0.906 | +12% |
| aerial-spheres | 318 | 11 | 0.898 | 0.989 | +10% |
Agents need three things in place before they can move: a model fitted to the site, operational tools to call, and data to learn from.
Superb AI already runs all three as one pipeline.
Superb AI developed and operates ZERO, its vision foundation model.
That gives wide control, from design to inference settings to retraining.
The MLOps platform and data engine are already connected.
Agents don't need new tools built to call.
Over 130 million industrial vision images refined to date. The only Korean company in NVIDIA's Physical AI ecosystem.
Superb AI is a vision intelligence company that turns industrial visual data into intelligence enterprises can use.
The Physical AI Data Factory, which builds the data robots learn from, and the Agentic Model Factory, which builds and runs the models that learn from it, connect as one pipeline.
Browse the topics below.
What is the Agentic Model Factory?
Agentic Model Factory is Superb AI's system in which agents design, optimize, and train vision AI models for production deployment. Agents take over the optimization, retraining, and model design work that people used to repeat at each site, while people set the criteria and review the results. It has three tracks: Agentic ZERO, Agentic MLOps, and Agentic Research.
What is Agentic ZERO?
Agentic ZERO is the track that optimizes ZERO, Superb AI's industry-specific Vision Foundation Model, for your data. Across multiple rounds, the agent tests different settings, including the examples and instructions given to ZERO, how it reads images, and whether to fine-tune. It keeps only changes that improve the score on held-out test data.
How is this different from the usual AI agent?
Conventional AI agents typically operate after deployment, using finished AI to get work done. Superb AI's agents work upstream: they design, optimize, and train the AI that will be deployed in production. What the industry calls agentic tooling helps people manage agents. Superb AI's agents, by contrast, manage the models themselves.
If agents retrain models, how do you keep control?
People set the decision criteria and the scope of automation first. The agent makes no decision outside them. Every adoption and rejection is logged with its evidence. Operators can stop the loop at any time and see what changed and why.
What does it mean for an agent to build a model?
Instead of a person selecting and tuning an off-the-shelf model, the agent designs, trains, and validates a model for the problem, grounded in published literature and code. Requirements and constraints such as target inference speed, memory limits, and unsupported operations are specified before design begins.
Is our data used for training?
Your data is used only to fit your model. Processing scope and retention period are specified in the contract. Isolated deployments, including on-premise, are supported.
Can we use it now?
Superb AI is looking for sites to validate with. Tell us what data you have and where your bottleneck is today, and you'll get a proposal for scope and process by track.
How can we help
We’ll show you how Superb AI can help :
Curate effective datasets to improve model accuracy by >15%
Build more accurate computer vision datasets 10x faster
Train, diagnose, deploy AI models with just a few clicks
The world’s leading AI teams trust us
"Superb Platform has been extremely useful and the AI-driven data curation makes it even more ideal. The ability to curate datasets and curate mislabels allows us to quickly and effortlessly build high-quality datasets."
Blaine Bateman, Chief Data Scientist
