OUR SERVICES
invariant.ai®
Lifecycle management of AI systems: managed services for the deployment, monitoring, compliance and governance of AI in critical processes.
Thanks to our experience of over 10 years bringing AI in production, we can offer our clients a managed service to create, deploy and monitor AI, ensuring absolute trust and reliability, compliance and a proper governance and the ability to generate better ROIs with a shorter time-to-value.
invariant.ai®, an end-to-end service:
Custom-tailored AI models and model fine-tuning, for all the situations not specifically supported by pre-trained models available in AI-go Studio and Detectiv.ai.
Invariant.ai® Runtime
A fast and reliable tool to execute AI models at the edge.
It allows to easily deploy AI models (and code) at edge (on multiple devices, like smart cams, industrial PCs, cloud VM), and integrate them into the client processes:
- ensure industrial-grade performance and reliability;
- high speed applications;
- timeout management;
- coordinate the output of multiple controls.
Work also in environment with limited or no connectivity.
Ensure IP protection of code and models with a flexible and configurable licensing.
Invariant.ai® Monitoring
A control room to access, monitor and configure all your AI processes.
Remote monitoring and alerting:
- monitor and receive alerts all your AI- processes from a single web-interface;
- benefits from proprietary AI metrics designed to detect model drifts;
- add business and process oriented KPIs, not only model related ones
- receive tailor-made periodic reports.
Remote configuration:
- verify the current configuration of all you machines;
- view the full change history with the ability to roll back.
Enable access to your data at multiple levels (client, OEM/integrator, Oròbix) to drive improvements. Digital twin for advanced reporting “beyond the edge”.
Invariant.ai® Governance and Compliance
Aimed in implementing the procedure for managing AI in production
AI Governance
Procedures to enable active learning and continuous model improvements:
- selective retention on data, based on the points more likely to improve a models;
- advices on the retraining policies and/or a managed retraining services managed by our data scientists;
- monitor the evolution of models over time, the reason that required each retraining, and the result “unit-like” tests before each model release.
AI Compliance
We work to ensure compliance with the standards and regulations imposed by the client’s industry:
- everything is tracked, from the datasets use for training and validation to the various model versions both at the cloud and edge level;
- access also to the full configuration history and who modified it;
- implement error-proof validation procedures to ensure that a model will continues to performance as expected.
invariant.ai® makes it possible to:
DEPLOY MODELS
Take models to production with guaranteed cycle times, providing complete observability and traceability of models and data through time.
MONITOR MODELS AND DATA
Identify anomalies in the data or during model execution and evaluate the actions needed to ensure model performance through time. Monitor performance and drift, and identify what data to collect to improve said performance following an active learning approach.
GUARANTEE COMPLIANCE
Manage risks deriving from the adoption of automated decision systems by integrating AI solutions into operational processes ensuring traceability and interpretability.
DEFINE ROAD TO PRODUCTION
Define the procedures for the adoption of AI systems in production, from validation to monitoring. Foster interoperability and enable cross-compatibility between different environments (e.g. Linux & Windows), by standardizing the messaging protocols between distributed functions, leveraging state-of-the-art stream-messaging technologies.
invariant.ai® provides answers to the most frequent questions surrounding the use of AI systems in mission-critical processes:
What model was being used for a given process at a given moment?
When did the model’s performance decline? What changed?
How did the data used to train and validate the model evolve?
What additional data should be gathered to improve performance?
What data are associated to high uncertainty?
Is the model performing in a way that meets my expectations and can it be used in mission-critical processes?
How can I validate a process managed by one or more AI subsystems?
Don’t say AI,
until you productionize
Oròbix takes AI to work in the real world.