Can computer vision run inside an iOS or Android app?
Yes. For in-app features such as document scanning, barcode and label reading or photo checks, we convert the model to Core ML for iOS and LiteRT (formerly TensorFlow Lite) for Android, so inference runs on the phone without sending images to a server. Larger models, or work that spans many cameras, run on edge hardware or in the cloud and the app calls them through an API. Accuracy is measured on your own images before launch rather than promised as a fixed figure in advance.
How much does Computer Vision development cost at Miracuves?
A ready-made platform from our catalogue starts from $2,199 and ships in 6 working days, at a fixed price when your scope matches the base product; vision features added to it are custom work. A custom Computer Vision MVP typically runs $8,000-$25,000 and takes 2-8 weeks depending on scope - the number of object classes, how much data must be collected and labelled, the number of cameras, edge hardware versus cloud, the accuracy target and the integrations decide where it lands. Ongoing Computer Vision development is available from $2,299/month for feature work, retraining and maintenance. Every quote is written before payment, with no surprise invoices after kickoff.
Do I need to provide a labelled dataset, or does Miracuves handle annotation?
Miracuves handles full annotation services for every project, and your domain experts review a sample of the labels. We use a tiered annotation pipeline: automated pre-labelling with our existing models, manual verification by trained annotators, and quality sampling by CV engineers. For standard use cases, we also have pre-labelled datasets available. You provide raw images or video; we deliver production-ready annotations.
Can Computer Vision models run on edge devices without the cloud?
Yes - edge deployment is our standard architecture for camera systems. Each model is exported to the runtime of its target: TensorRT on NVIDIA Jetson, ONNX Runtime or OpenVINO on industrial PCs and Intel hardware, and LiteRT on Raspberry Pi-class boards and phones. Inference runs on the device, and latency is benchmarked on your hardware before we commit to a frame rate. Cloud connectivity is used only for model updates, dashboard syncing and remote monitoring - never required for real-time inference.
Computer Vision vs AutoML Vision APIs - when should I choose custom CV development?
AutoML APIs work for generic use cases with standard object classes and non-critical accuracy requirements. Choose custom Computer Vision development when you need: industry-specific defect categories not in any pre-built API, low-latency inference on edge hardware, full control over model architecture and training data, or an accuracy target that a generic API does not reach on your own images. Miracuves provides a free scoping call to determine which path fits your use case.
How does Miracuves handle model retraining when production data reveals new edge cases?
Every production deployment includes an Active Learning pipeline. When the monitoring dashboard detects accuracy drift below threshold, or when operators flag false positives/negatives, the system automatically saves these edge-case images. Our CV team reviews, annotates, and retrains the model - typically within 48 hours. The updated model passes through the full QA gate before deployment.
What hardware does Miracuves recommend for running CV models in production?
For industrial vision on the edge we usually recommend an NVIDIA Jetson Orin module, sized to the model and the number of camera streams; for lightweight applications, a Raspberry Pi 5 with an add-on AI accelerator can be enough. For cloud inference, we deploy on GPU instances sized to your throughput, with auto-scaling. Miracuves recommends hardware from your throughput, latency and power requirements, and benchmarks the model on it before you buy in volume - no over-provisioned infrastructure.
Does Miracuves handle data privacy and compliance for regulated industries?
Yes. We design data handling for regulated industries from day one and build to the controls those rules require; certification and compliance sign-off remain your organization's audit. For healthcare, we build to HIPAA's safeguards, with access controls, audit logs and de-identification of images before they leave your network. For manufacturing, inference can stay entirely on-premise, so no image data leaves the factory floor. For security and retail cameras, we support retention limits and face blurring for people who have not consented, as GDPR-style rules require.
What is the typical timeline and process for a Computer Vision project?
A custom Computer Vision engagement typically runs 2-8 weeks in five phases: (1) data audit and labelling, (2) model training and validation on your images, (3) optimization and testing on the target hardware, (4) integration with your systems, and (5) deployment and monitoring setup. Larger scopes - many camera sites, large labelling volumes or novel model architectures - are quoted in writing before work starts.
Can Miracuves provide dedicated computer vision developers?
Yes, as a Miracuves team rather than individual contractors. On the Ongoing CV Development retainer, from $2,299/month, a dedicated team works on your product with weekly sprint demos: new object classes, retraining on fresh images, new camera sites and releases. You talk to the engineers directly, the team scales up or down with the work, and all code and trained models stay 100% yours. For a defined first project, a fixed-scope custom build is usually the better start.
Can you use our existing cameras or CCTV system?
Often, yes. We pull streams from IP cameras over RTSP or through your video management system, and start by reviewing sample footage from each camera. Existing CCTV usually works for counting people or vehicles, occupancy and presence detection. It is often too compressed, too distant or too dark for fine inspection or reading small text; in those cases we say where a dedicated camera or better lighting is needed before any model work is quoted.
Can you improve a computer vision model another team built?
Yes. We start with a written audit: the training data and how it was labelled, the model and its export for the target hardware, and a fresh accuracy test on a held-out set of your current images - many models degrade because products, packaging or lighting changed after training. The plan that follows may be relabelling, retraining, a newer pretrained model or moving inference to different hardware. The work is quoted as custom Computer Vision development before it starts.
