Dulce™ — Batch Image Classification for Windows®
Run supported image classifiers on Windows® — without building the operator application
Run recurring batches of prepared images with supported ONNX classifiers using CPU, DirectML, or optional NVIDIA® CUDA acceleration.
Beta 1 — 60-day evaluation for Windows® 11
PRODUCT
Run the classifier without building the surrounding operator application
Dulce™ provides a Windows® application for running supported, separately supplied ONNX image classifiers against prepared image batches. A model provider, engineering team, or integrator can configure the classifier's required execution and preprocessing settings for use by a downstream Windows® user without requiring a purpose-built inference application, Linux server, cloud inference service, container stack, or Python environment solely to run recurring batches.
Keep images under your control
Run supported batch classification locally on Windows® hardware without a required cloud inference path.
Put your workstation to work
Use compatible CPU, DirectML, and optional NVIDIA® CUDA devices already available in the workstation
Review results after the run
Retain structured batch inference results and run evidence for comparison, investigation, and later analysis
USE CASES
Where Dulce™ fits
You need someone else to run the classifier — Use Dulce™ when a model owner, engineering team, or integrator needs to hand a supported classifier and its known preparation recipe to another Windows® user who will run recurring image batches.
You already have images and a supported model — Use Dulce™ when you already have a prepared image set and a supported discrete-class ONNX model and need a controlled Windows® batch-classification workflow.
You want to keep inference local — Run image classification on customer-controlled Windows® hardware without requiring a cloud inference path or Internet-connected runtime.
You want to replace ad hoc batch tooling — Move from scripts or manually assembled inference workflows to a structured application that records per-image results, run-level information, and execution evidence.
You want reviewable results after the run — Use Dulce™ when classification results need to remain available for later review, documentation, investigation, or downstream analysis rather than existing only as transient runtime output.
A supported ONNX model is only part of the execution contract. The model owner must also know the class labels and order, required image preparation, and output interpretation used by the classifier.Dulce™ works with previously captured image files. It does not provide model training, camera control, PLC integration, object detection, segmentation, metrology, or real-time production-line control.
Manufacturing is one application
Typical uses include evaluating classification models against captured production images, re-examining historical image sets, comparing model behavior across datasets, and performing offline analysis of manufacturing imagery. These are examples of the broader Dulce™ workflow: running supported classifiers against prepared image batches when repeatable Windows® execution is needed.
WINDOWS® + LOCAL
Batch inference without a separate AI infrastructure stack
Stay local
Images and inference workloads remain on the workstation rather than requiring a cloud inference service. Runtime operation does not require Internet or cloud connectivity.
Use existing hardware
CPU, DirectML, and optional NVIDIA CUDA execution allow Dulce™ to use hardware already available in the workstation.
Use Windows® directly
Dulce™ runs as a Windows® desktop application without requiring operators to work with Linux, containers, or Python environments.
Keep the workflow simple
Models, image sets, execution settings, and batch results are managed through the application rather than assembled from command-line tools and scripts.
RESULTS
Reviewable batch inference output
Dulce™ records classification results and run-level information so completed batch runs can be reviewed, documented, and analyzed after processing.
Per-image results
Review classification results for the images processed during the run instead of relying only on transient runtime output.
Run-level summary
See run-level information including image counts, timestamps, throughput, and the locations of generated result artifacts.
Execution evidence
Dulce™ records per-device execution telemetry together with the model SHA-256 and, when applicable, the label-file SHA-256.
Results that remain useful after the run
Generated results can support later review, investigation, documentation, or downstream analysis using the artifacts produced by Dulce™.
Examples from a completed run:
Completed run summary
Per-image classification results
Machine-readable run summary
System requirements
Dulce™ Beta 1 is designed for 64-bit Windows® 11 workstations
Windows® 11
64-bit Windows® 11 desktop or workstation
Execution hardware
CPU, DirectML, and optional NVIDIA® CUDA acceleration
Storage and memory
About 1.01 GB installed, with approximately 1.5–1.6 GB typically observed during batch inference
OPTIONAL NVIDIA® CUDA
CUDA acceleration
CUDA acceleration is optional. Dulce™ can run using CPU or DirectML without NVIDIA® CUDA dependencies.*
| CUDA 12 | cuDNN 9 |
|---|---|
| A compatible NVIDIA® GPU, NVIDIA driver, and CUDA 12 runtime are required | cuDNN 9 runtime required and configured in Dulce Settings |
| CUDA downloads | cuDNN downloads |
* Dulce™ does not install or update NVIDIA CUDA dependencies
CUDA optional setup
DOCUMENTATION
User Guide
The Dulce™ Beta 1 User Guide covers installation, supported models and image inputs, CPU/DirectML/CUDA execution, CUDA requirements, batch operation, persistent data, and uninstall behavior
Beta 1
Apply for the Dulce™ Beta
Who Beta 1 is for?Beta 1 is intended for qualified U.S. businesses evaluating Dulce™ with a real supported image-classification workflow.
A strong Beta candidate has:
a U.S. business evaluating Dulce™ for a real image-classification workflow
a supported discrete-class ONNX classifier
the classifier's class labels and class order
authoritative image-preparation and output settings
prepared image files the business is authorized to process
a Windows® 11 x64 workstation
the Windows 11 workstation must be able to sign in to the Microsoft Store with a personal Microsoft account, because a work or school account cannot install the current private listing
a recurring batch-classification need or a concrete upcoming classifier handoff; and
for handoff evaluation, a downstream Windows® user other than the classifier's developer who can run the workflow.
Dulce™ Beta 1 works with previously captured image files. It does not connect to cameras, PLCs, or production-line controls, and it does not provide detection, segmentation, measurement, or automatic reject decisions.Applicants use a business email address.The Windows® 11 x64 workstation used for the evaluation must be able to sign in to the Microsoft Store with a personal Microsoft account. A work or school login cannot install the current private listing.Beta 1 is a 60-day evaluation on that computer.CUDA is optional. CPU and DirectML do not require NVIDIA CUDA.I confirm that I am applying on behalf of a business, am using a business email address, have access to Windows® 11, and have a Microsoft® account for Microsoft® Store access.Apply for Beta 1
Support & Beta Feedback
Contact Dulce™
Questions about Dulce™ Beta 1, installation, CUDA configuration, or application behavior can be submitted using the contact form.Please include a brief description of what you were doing and what happened. We may contact you for additional technical details if needed.Beta feedback is encouraged. Use this form for support questions, problems, successful use cases, documentation feedback, or feature suggestions.Please do not include proprietary, confidential, or sensitive information in the form. If diagnostic files or other information would be useful, we will request them separately.See Privacy for information about how submitted information is handled
PRIVACY
Privacy and data practices
Dulce™ application
Dulce™ is designed for local, on-premises operation. The Dulce™ application does not collect, track, store, or transmit user data or application telemetry.
Website
This website may process limited technical information needed to deliver and secure the site, such as IP address, browser or device information, and server log data. Service providers used to operate the website may process this information on our behalf.If you contact Sweet Software Practices Corporation through this website, we receive the information you choose to provide, such as your name, company name, email address, and message. We use that information only to respond to your inquiry and provide requested support or information.
Beta feedback
Information voluntarily provided as Beta feedback may be retained and used to evaluate and improve Dulce™.
Privacy questions
Use the Support contact form
SWEET SOFTWARE PRACTICES CORPORATION
Dulce™
Windows® batch inference for manufacturing
Oshkosh, WI USA
© 2026 Sweet Software Practices Corporation
Support · Privacy · User Guide