AI Studio is eSmart Systems' computer vision AI platform. It lets you build, test, and deploy custom visual AI models, and chain them into multi-step inspection pipelines, without any machine learning expertise. It's accessible via a web interface at ai.esmartsystems.com and via API.
The current release is a Research Preview, a fully functional early version of the platform made available so that users can explore and test capabilities while we continue to refine and expand the product. You may encounter new or changed features or changes to user experience with each visit, and we welcome your feedback.
You can register at ai.esmartsystems.com using SSO via your Microsoft or Google-hosted email account. The Free tier requires no payment.
Check out our demo video on the About Adaptive AI page.
The Free tier gives you access to the Model Builder to create and test Adaptive AI classifiers and the Pipeline Builder to create and test AI model pipelines, both with limited usage.
The Enterprise tier unlocks full platform access, including the Model Garden with eSmart Systems' pre-built model suite and unlimited usage limits, and API access.
See the Pricing page for current tier details.
The Model Builder lets you create custom AI models from a small number of example images, using few-shot techniques. No traditional model training, labeling tools, or GPU infrastructure required.
The Model Builder supports two parallel few-shot technologies, each covered in its own section below:
They solve different problems. Pick by what you need the model to do:
The two are commonly chained in a pipeline: FSOD locates the objects, then an Adaptive AI classifier labels each detected crop. The Pipeline Builder makes that straightforward.
Yes. Both Adaptive AI classifiers and FSOD detectors support real-time tuning. Add new examples, remove poor ones, or adjust class definitions, and the model picks up the changes immediately. No retraining step. Open the model in the Model Builder, edit the examples, press Save, and the model is updated.
Models you create are visible under My Models. Once you save a model, it is automatically shared with your team Model Garden as well, with the same behavior for both model types.
An Adaptive AI classifier assigns a class label to a whole image or a crop, for example "is this insulator damaged?" or "what type of pole is this?". It's built on eSmart Systems' patent-pending Adaptive AI technology, requires only a small number of example images per class, and is available to every Model Builder user.
Around 5 example images per class is the minimum for reliable results. More examples generally improve accuracy, and variety in viewpoint, lighting, and background helps the model generalize.
Upload example images for each class. No bounding boxes, no annotation step. The classifier learns the class boundary directly from the example images you assign per class.
Fast and Accurate are backbone options you choose when creating an Adaptive AI classifier. Fast is best for less complex problems where inference speed and iteration in development are more important than performance. Accurate uses our most capable backbone, which makes it suitable for harder visual tasks at the cost of higher latency.
Few-Shot Object Detection finds and locates objects in an image based on a small number of bounding-box examples. The model returns a bounding box and a confidence score for every instance it detects, with no traditional training or large labeled datasets required. FSOD is an account-scoped capability. If you don't see it in the model-type picker, your account hasn't been granted access yet. Reach out via the Contact Us page or talk to your eSmart Systems account executive to enable it.
2 to 10 bounding-box examples per object class is a good starting point, drawn across one or more uploaded images. During development, switch to Test mode after a couple of examples to see how the model performs, then add more or refine existing ones based on what you see. Variety in viewpoint, lighting, and background helps the model generalize.
You upload full images and draw bounding boxes around each target object directly in the Model Builder's Annotate mode. Those bounding boxes are the examples the model learns from. There is no separate annotation tool to learn, no labeled-dataset step, and no need to pre-crop your images.
Alongside your bounding-box examples, each FSOD class takes a short text prompt that describes what to find. A concrete noun phrase usually works best, like "rusty bolt", "cracked insulator", or "yellow car". Two to four words is a sensible target.
Experiment with multiple prompts during development. Small wording changes (more specific, less specific, different adjectives) can have a big impact on what the model picks up. Compare results in Test mode and keep the prompt that gives the best precision and recall balance.
Annotate mode lets you draw or correct bounding boxes on full images to add new examples or refine class definitions.
Test mode runs the model against an image so you can inspect detections, tune the confidence threshold, and validate behavior before pushing it into production.
Every FSOD detection comes with a confidence score between 0.0 and 1.0. The threshold is the cutoff below which detections are discarded. Lower it to surface more (but less certain) detections; raise it for higher precision. The threshold is set per model and applied at inference time, including when the model is called via API or as part of a pipeline.
A Dataset is a saved, reusable collection of images. Datasets can be used as input when testing models in the Model Tester, and as training data when building models in the Model Builder.
Yes. Toggle to the Team Datasets in the Datasets section to see collections shared across your organization. These can be used directly in the Model Tester or Model Builder. You are able to edit Team datasets, including naming, add/remove images, and give the dataset a description.
Uploaded content is screened using Microsoft Azure AI Content Safety. Inappropriate or harmful imagery is blocked. All users must comply with eSmart Systems' Terms of Use Agreement, which is updated from time to time. Accounts found to be in violation will be suspended.
The Model Garden is the library where your AI models and pipelines are hosted. Once you have created a model or a pipeline, they will be available for inference/operation in the Model Garden. The Model Garden also exposes the APIs for your models and pipelines.
For relevant enterprise customers, the Model Garden will also be pre-populated with production-ready AI models: detectors, classifiers, and pre-built pipelines. This includes eSmart Systems' own models trained on utility infrastructure (transmission, distribution, substations) as well as general-purpose models and any models your team has created and shared.
The Model Garden includes:
You can filter by type (Detectors, Classifiers, Adaptive), model creator (team, personal) and use case.
In the Model Garden, use the Team filter to see models created and shared by others in your organization. These are available to use in the Model Tester and Pipeline Builder.
Yes. Click Test Model (or Test Pipeline for pipeline entries) on any model card to open it directly in the Model Tester.
Yes. Model Garden supports BYOM (Bring Your Own Model) for Enterprise customers. Additional framework support is on the roadmap.
The Pipeline Builder lets you compose multi-step visual AI workflows by chaining models together. For example: detect a component → classify its condition → measure a value. Pipelines are built using a visual drag-and-drop editor and can be deployed as API endpoints.
Yes. The left panel in the Pipeline Builder shows all available models, including eSmart Systems' pre-built models, as well as Adaptive AI classifiers and FSOD detectors created by you or your team.
Note: eSmart Systems' models are not available as part of the Free tier.
Once you are ready to deploy a pipeline, open the pipeline in the Model Garden, which will open the pipeline in the Model Tester. Here you can open the API Access drop down and get the details for the API endpoint, content type, and body, with an example to copy.
Note: you need to create an API key prior for access to API call.
You can access AI Studio API documentation here, and MCP Integration documentation here.
Yes. Use the Load button in the Pipeline toolbar to load a saved pipeline.
The Model Tester is a testing environment where you can run individual models or full pipelines against images, either uploaded from your local machine or drawn from a saved Dataset. Results are displayed in real time.
Yes. The model selector includes pipelines as well as individual models. Select a pipeline from the dropdown to test it end-to-end.
Once you are ready to deploy, open the model or pipeline in the Model Garden, which will open it in the Model Tester. Here you can open the API Access drop down and get the details for the API endpoint, content type, and body, with an example to copy.
Note: you need to create an API key prior for access to API call.
Yes. Every model and pipeline in AI Studio exposes an API endpoint, allowing you to integrate AI Studio capabilities into your own applications, workflows, or automation pipelines.
You can access AI Studio API documentation here, and MCP Integration documentation here.
API calls are handled the same way as the use of AI Studio via the website, in that all count towards the consumption of Vision Transactions (VTs). More information on pricing and plans available on the pricing page.
When you're part of an organization on AI Studio, you can view and use models, pipelines, and datasets that your teammates have shared under the Team tabs in Model Builder, Model Garden, and Datasets.
Whenever you create and save a model, pipeline or dataset, this is automatically added to your Team tabs as well.
In the initial Research Preview this is not supported but it will be added in a near term update.
One image processed by one model counts as a Visual Transaction (VT). Different model types have different costs per image:
In a pipeline, each model invocation is counted separately. If a detector finds 5 crops and a downstream classifier runs on each crop, that classifier node charges 5 VTs. Skipped or conditional nodes that don't actually run cost nothing.
Enterprise customers have unlimited VTs and can monitor team consumption in the Usage tab on the Enterprise page.
Failed inferences are not charged. If a model errors out (inference service unavailable, network failure, invalid input, etc.), no VTs are deducted. In a pipeline, only the nodes that complete successfully are charged. If step 3 of 5 fails, you pay for steps 1 and 2 only.
Empty or low-confidence results that complete successfully are still charged, because the inference itself ran. Finding nothing is a valid result, and the compute cost is the same.
Submit a request via our Contact Us page and a member of our team will be in touch. To compare plans, visit our Pricing page.
You can review the terms of use of AI Studio at ai.esmartsystems.com/terms. If you are an Enterprise customer, your negotiated license terms will apply. Check with your eSmart Systems account executive if you have any questions.
For any queries, feedback or support, visit the Contact Us page.
If you are an existing eSmart Systems' Grid Vision customer, continue to use the support portal and email support at:
AI Studio is a product of eSmart Systems. This FAQ reflects the Research Preview release and will be updated as the platform evolves.