Does Miracuves build custom ML models or use pre-built ones?
Miracuves builds custom models tailored to your data and use case. While we leverage proven frameworks like TensorFlow, PyTorch, and scikit-learn, every model is trained on your specific data with your specific success metrics. We do not deploy generic pre-built models as production solutions unless your use case genuinely calls for it.
How much does machine learning development cost at Miracuves?
A scoped ML module - data pipeline, trained model, inference API and monitoring dashboard - starts from $3,699 and takes 4-8 weeks; the price stays fixed when the use case matches a proven template such as recommendation, churn or forecasting. A custom ML platform typically runs $18,000-$80,000 and takes 2-8 weeks, with larger scopes quoted in writing before work starts; data volume and quality, classical versus deep learning, real-time inference, the number of data sources, retraining and A/B testing scope, compliance needs and edge versus cloud deployment decide where it lands. Ongoing ML work is available from $2,299/month for feature work and maintenance. Every quote is written before payment, with no surprise invoices after kickoff.
What data do I need to provide for an ML project?
The data requirements vary by use case, but typically include historical records relevant to the prediction target - transaction logs for fraud detection, sensor readings for predictive maintenance, user behavior data for recommendation engines. Miracuves helps assess your data readiness in the discovery phase and can recommend data collection strategies if gaps exist.
How long does it take to deploy an ML solution?
A scoped ML solution - data pipeline, model training, API deployment, and monitoring dashboard - ships in 4-8 weeks depending on data complexity and model requirements. Custom builds with unique architectures take 2-8 weeks, with larger scopes quoted in writing before work starts. All timelines include testing, validation, and deployment. Timelines are stated in writing before any payment.
What is included in the ML monitoring and maintenance?
Every ML delivery includes a monitoring dashboard (Grafana) tracking model accuracy, prediction latency, data drift, and system health. Automated alerts trigger when performance degrades. Miracuves includes 60 days of post-deployment support. Model retraining and ongoing MLOps are available through our monthly retainer model.
Can Miracuves deploy ML models on edge devices?
Yes. We support edge deployment using TensorFlow Lite and ONNX Runtime for devices with limited compute. Computer vision models, in particular, are optimized for edge inference with quantization and pruning. For cloud deployment, models run as containerized microservices on Kubernetes with auto-scaling.
Do I need a large dataset to work with Miracuves?
Not necessarily. Miracuves works with datasets of all sizes. For smaller datasets, we use classical ML approaches (scikit-learn) with careful cross-validation to avoid overfitting. For larger datasets, we leverage deep learning with GPU acceleration. If data is truly insufficient, we will recommend starting with data collection before model development.
How does Miracuves handle model versioning and reproducibility?
Every model experiment is tracked in MLflow - parameters, metrics, and artifacts are logged and versioned. The full pipeline is defined as code (Kubeflow), meaning any previous version can be fully reproduced. Model registry maintains a history of deployed versions with rollback capability. This is a non-negotiable standard on every project.
What happens if model accuracy degrades after deployment?
Monitoring dashboards track accuracy and drift continuously. If degradation is detected, Miracuves investigates root cause - data drift, concept drift, or infrastructure issues. Under the 60-day post-deployment support window, corrective retraining is included. For ongoing needs, our monthly retainer covers model retraining, pipeline updates, and continuous improvement.
Can Miracuves add fraud detection to a Gojek-style super app?
Yes. Our ready-made Gojek clone ships as a multi-service super app in 6 working days. Fraud scoring on top of it is custom ML work: we log the signals the platform already captures - devices, trips, payments, promo use - build features from them, and deploy a scoring API with a review dashboard once enough labeled cases exist. Until then, rules cover the known patterns. The ML work is quoted separately before it starts.
Is Miracuves machine learning development tied to one framework?
No. We pick the framework after we have seen the data. Tabular problems such as fraud, churn and demand forecasting usually run on scikit-learn, XGBoost or LightGBM; images, audio and long sequences go to PyTorch or TensorFlow. The model is served behind the same kind of API either way, so your application does not depend on the choice. If you already run TensorFlow in production and want to stay on it, our TensorFlow development service is built for that.
Can you take over a model our team built in a notebook?
Yes, and it is one of the most common starts. We first reproduce the notebook's result from raw data, which often exposes leakage or a split that flattered the model. Then we move the feature logic and training into a versioned pipeline, register the model in MLflow, wrap it in an inference API and add monitoring. Your data scientists keep ownership of the modeling choices; the work is quoted before it begins.
