ModelNova Fusion Studio: An Ecosystem for Building Edge AI Applications
embedUR has been focused on making edge AI development more accessible for engineers and product teams. Over time, we built ModelNova, a library of ready-to-use Edge AI models and datasets, and Fusion Studio, a desktop environment for training, labeling, and customizing those models. Together, they form a workflow that removes much of the friction from developing embedded AI systems. Let’s explore inside ModelNova and Fusion Studio and the real-world applications they make possible.
Inside ModelNova
ModelNova is a complete resource for Edge AI development. Think of it as a Model Zoo and desktop IDE rolled into one to make it easier and faster for engineers to prototype, train, and deploy AI on edge devices without starting from scratch. ModelNova gives you:
- Pre-trained, benchmarked models
- Curated datasets
- Blueprints and silicon benchmarks
- Documentation and community support
You can download a model, test it, and see how it performs on hardware platforms like Raspberry Pi. Then, if you need a custom function, Fusion Studio makes it easy to adapt, retrain, and optimize the model locally. With ModelNova, ideas turn into prototypes quickly. Instead of spending months creating a model from scratch, you can build a proof of concept in weeks and focus on other important features of your application.
Edge AI Models Inside ModelNova
Here’s a look at some of the models inside ModelNova.
ResNet50 – Image Classification: ResNet50 is a proven benchmark for image recognition. It is versatile and efficient for tasks like household or food classification. Its architecture balances accuracy with performance, making it suitable for devices with limited compute.
ResNet101 – Advanced Image Classification: ResNet101 is a deeper version of ResNet50. It provides higher accuracy for more complex image datasets. It’s ideal when finer-grained distinctions are required.
MnasNet05 – Lightweight Classification: Designed for ultra-low-power devices, MnasNet05 is small and fast, perfect for embedded vision tasks.
Textify OCR – Text Recognition: Textify converts images of text into digital content, supporting applications that need optical character recognition on edge devices.
RoBERTa Base – Question Answering: A compact language model for understanding and generating text, optimized for edge deployment.
MicroSpeech LSTM – Speech Recognition: A lightweight audio model for recognizing simple voice commands. Its small size makes it ideal for devices like smart speakers or industrial voice interfaces.
Featured Datasets Inside ModelNova
ModelNova also provides a curated set of datasets designed to cover common edge AI applications. Let’s explore some of the datasets:
WIDERFace – Face Detection: A large-scale dataset with over 32,000 annotated faces, WIDERFace is ideal for training facial recognition and detection models. It’s used in applications like access control, driver monitoring, and security cameras.
CIFAR-100 – General Image Classification: CIFAR-100 contains over 60,000 images across 100 categories. This dataset is perfect for testing image classification models like ResNet50 or ResNet101.
Wake Vision – Domain-Specific Object Detection: Wake Vision is a large-scale dataset with over 6 million files designed for edge object detection tasks. It’s tailored for applications like obstacle detection in robotics or environment monitoring.
Fusion Studio: Training, Annotation, and Labeling on Desktop
Fusion Studio is a desktop IDE designed specifically for edge AI development. It gives engineers full control over the training and customization of their models without relying on cloud infrastructure. Fusion Studio gives you access to:
Local Model Training
Fusion Studio lets you train your models directly on your desktop. You can experiment with hyperparameters, try different training strategies, and validate results immediately. This approach avoids cloud compute costs and reduces dependency on external infrastructure.
Integrated Dataset Annotation and Labeling
You don’t have to juggle multiple tools. Fusion Studio includes built-in modules for annotating and labeling datasets. Whether you’re marking bounding boxes for object detection, segmenting images, or labeling audio clips, all steps happen locally. This keeps sensitive data on your machine while speeding up preparation for training.
Rapid Iteration
Iteration is faster because training, labeling, and testing happen in the same environment. You can quickly adjust datasets, retrain models, benchmark performance, and optimize without exporting files or switching platforms.
Model Benchmarking and Optimization
Fusion Studio provides performance metrics like inference speed, model size, and accuracy, helping you select the best model for your target edge device. You can also optimize models (pruning, quantizing, or adjusting layers) before deploying them to production hardware.
Seamless Deployment
After training and optimization, Fusion Studio simplifies the process of porting models to your preferred edge AI hardware. There are custom compilers and deployment features that will ensure that your models run efficiently and reliably on your hardware.
Applications at the Edge
i) Quality Control in Manufacturing: A ResNet-based image classification model can be used with cameras on production lines to catch defects in real time. Because the processing happens locally, manufacturers avoid cloud delays and keep sensitive data in-house.
ii) Voice Interfaces in Healthcare: A lightweight speech model like MicroSpeech LSTM enables hands-free control of bedside devices or monitoring equipment. Patients can interact without pressing buttons, and hospitals gain reliable systems that work even offline.
iii) Smart Farming Tools: Edge vision models help farmers track crop health, detect pests, or monitor soil conditions using small drones and sensors. These systems run in the field without constant connectivity.
iv) Retail and Inventory Management: OCR models such as Textify can scan receipts, labels, and stock automatically at kiosks or handheld devices. Shops can also use vision models to track items on shelves and detect when stock is running low.
v) Driver and Passenger Safety: Models trained on datasets like WIDERFace and pose estimation can run inside vehicles to detect drowsiness, distraction, or unsafe movements. This keeps drivers alert and helps prevent accidents without relying on internet access.
From Idea to Deployment with Model Nova and Fusion Studio
Model Nova is designed to remove the heavy lift that usually comes with starting an edge AI project. Instead of spending months sourcing datasets or building models from scratch, teams can draw from a curated library of pre-trained models, tested datasets, and silicon benchmarks. This foundation makes it possible to choose the right architecture and hardware combination early, cutting down trial-and-error and wasted cycles.
Fusion Studio then takes over as the desktop environment where all the actual engineering work happens. You can annotate and label datasets directly inside the platform, train models locally on your workstation, and experiment with hyperparameters without relying on cloud infrastructure. Because training, optimization, and benchmarking are built into the same workflow, you see results immediately: model accuracy, inference speed, and memory footprint against your chosen device.
The process is faster because iteration doesn’t leave the desktop, and it’s safer because sensitive data never leaves your environment. Together, Model Nova and Fusion Studio eliminate scattered tools and fragmented processes, giving teams one integrated path from concept to a working proof-of-concept.
If you’re working on an edge AI project, you can get to MVP faster with Model Nova Fusion Studio – the desktop IDE for building Edge AI applications.
