Most companies struggle to scale their artificial intelligence efforts. From proof of concept to deploying multiple projects in production that are adopted by clients or internal end-users, that generate return on investment, the road to success is a real challenge. Compliance, data security and privacy, integration of the AI initiatives with the existing software ecosystem, managing change with clients and end-users as well as scaling the initiatives beyond the data-scientists are amongst the obstacles that need to be overcome to finally generate a substantial ROI. Solutions to these obstacles need therefore to be thought of and adopted from the ground up. Adopting the right tooling can significantly help as well to accelerate, scale to multiple users, business units and geographies while integrating the constraints of security and compliance teams from the start.
Auto-ml
The development of open-source machine-learning frameworks such as Pytorch, Tensorflow or Scikit-learn that have democratized machine-learning to savvy developers and to a large number of apprentice and experienced data-scientists. With these frameworks, they could focus on using machine-learning algorithms and understanding better the data that will be used during modeling rather than understanding the complicated bits required for crafting and training efficiently their algorithms and have got access to a wide range of algorithms for different purposes such as classification, regression, natural language understanding or time-series modelling.
However, most of these algorithms are often parametric algorithms and it is generally more an art than a science to find the right set of parameters to find the optimal performances. Moreover, ‘data scientists can’t get free lunch’ (There is No Free Lunch in Data Science – KDnuggets) and therefore cannot rely on a single algorithm to solve all the problems they are facing. Not only should they understand the data, but they should also assess different algorithms to find one that is optimal on the given dataset. This can be a time-consuming task and that’s why companies such as DreamQuark have focused on helping data-scientists accelerate this search for the most optimal algorithm for their task and data. With this feature, data-scientists can assess hundreds of combinations much faster until our software can provide the best selection of models with key information such as performance metrics and explanations to choose which one to deploy. Thanks to auto-scaling capabilities, it can take a few minutes in the cloud to get the first models and it takes in average a month to build and deploy a first model in production (after all check and validations are made).
Moreover, using these frameworks require to know coding and to learn these frameworks. Even if they simplify the access for developers and data-scientists, the learning curve can be steep for business analysts without these skills which prevents them from having such a powerful tool to their list. Auto-ml in combination with intuitive interfaces facilitates the use of these algorithms and provides business analysts with the capabilities to use machine-learning to solve difficult business problems and explore data to get deeper insights on the business issues that their parent company is facing.
Multi-tenant
Scaling AI in a large company, present in multiple geographies with multiple business-units (and required Chinese walls), is not easy. Data governance, regulatory compliance and security may prevent that the data is not physically or logically segregated. A team in a region or a business unit may be required to have their data and their initiatives in an environment that could not be accessible by other teams from other geographies and business units. In that environment, however, data can be freely shared across members to build and deploy the initiatives.
Companies that want to adopt AI and scale their initiatives should therefore be warrant of adopting multi-tenant architectures or to rely on software that are providing this guarantee as they are built on such principles.
DreamQuark offers such an architecture and companies using DreamQuark AI operationalization software solutions have set them up with multi-regions architectures that guarantee the privacy of sensitive data across regions while enabling a freedom of development and agility within these regions.
Security
With stricter regulations around data-privacy across the globe, an increasing number of digital risks and data leaks, but at the same times business imperatives that demand agility, the recourse to the cloud and internal collaboration, it has become business critical to embed security and privacy at the core of data and machine-learning initiatives and to build or adopt solutions that are secured and private by design.
Security and machine-learning may seem contradictory as data-scientists need to explore data to be able to build powerful algorithms. Moving sensitive data to the cloud may seem like an even bolder move for non-US companies in particular with the ‘Cloud Act’ that gives federal agencies the right to access to personal data that are stored by US companies.
In front of these challenges, differential privacy and data-obfuscation are the Chief Security Officer best ally to meet the security expectations of its company customers while providing at the same time a way to meet the expectations of business units that want to leverage advanced data-analytics and machine-learning to meet their business objectives.
DreamQuark provides such functionalities in a transparent and intuitive way to enable data-scientists, business analysts and the end-users of the algorithms to leverage advanced ML algorithms in a secured and private set-up that enables for example companies to move sensitive data across geographies or move data to the cloud without affecting the performance of the algorithms and the usability of the software. DreamQuark provides functionalities for dependencies and vulnerability management as well as data and state backup. Finally, we provide SSO and fine grained user role management to restrain or allow the access to datasets, models or deployed models.
Hybrid and Multicloud
As companies are accelerating their migration to the cloud worldwide, an increasing number of companies are adopting multicloud models to avoid vendor lock-in and benefit from the combination of the strengths of different cloud providers. Other companies, particularly industries with sensitive data (such as the financial services industry) are adopting hybrid models to leverage their existing infrastructure while migrating their new software applications, or less sensitive applications, in the cloud, or to benefit (and learn) from the wealth of innovations that cloud providers are offering.
Companies may then be in a situation where some business units and/or geographies have adopted a cloud provider while others have adopted another provider. The other situation is when personal data is involved. Workloads that do not involved personal data (such as processing market data for portfolio optimization or risk management, or when anonymized data is involved) are moved to the cloud while workload involving personal sensitive data stay on premise.
In order to scale, companies need a way to create consistency across all their initiatives, across all their infrastructures, with common policies and standardized processes. They also need a way to create for example an initiative that leverages anonymized data in the cloud that is then deployed on premise or on another cloud (because development teams and deployment teams are leveraging different cloud providers).
DreamQuark leverages Kubernetes to orchestrate the containers that are used to deploy all the services behind its software suite. Kubernetes enables us to deploy our software across multiple environments either in the cloud or on premise, as well as distributions of Kubernetes. We support AWS, GCP, Microsoft Cloud or IBM Cloud. Our models can be trained in the cloud and deployed elsewhere. Moreover, we also enable our customers to benefit from advanced backup and disaster recovery functionalities through a partnership with NetApp, as well as all the stack that NetApp provides on top on ONTAP to help companies manage their hybrid cloud initiatives.
Collaboration
Companies that successfully deploy AI adopt a collaborative approach leveraging teams of diverse individuals and skills (with compliance, business, data engineers, data-scientists, business analysts and end-users). They also leverage data-sharing to allow several teams to work on the same datasets for different purposes (for example a CRM dataset could be used to help marketing prepare future campaigns, business development teams identify their next best actions and customer service teams identify customers the most likely to churn).
In these conditions it is key to ease data-sharing across the organization and to provide tools that are adapted for the different roles involved in the deployment of the AI initiatives. The end-users have different needs in comparison of the compliance teams, and data-engineers have different needs in comparison of the persons that need to validate that the model is adapted and its ROI potential sufficiently high.
DreamQuark provides tools for such a diverse team and has put the emphasis on collaboration. Hence, data engineers can leverage our data-pipeline and customer data platform to facilitate the aggregation of internal and external data and implement scripts to improve data quality. Through Brain it is possible to share datasets between several persons (with or without a right to download them) as well as models and APIs. Several persons can work on the same dataset or on the same project or can work in parallel on several projects leveraging the same data before combining these projects in a single workflow. It is also possible to give temporary access to people that come from outside your organization and remove their rights without losing their work. We also provide tools for the persons that will be in charge of validating and monitoring the models once they will be deployed.
Through the interfaces that we provide, the least expert individuals can build powerful models but it is also possible to interact with the different functionalities through a SDK to automate part or the whole process (from data upload to model monitoring)
Last but not least, a model is creating value if it is used by end-users. It takes only a click to deploy this model to make it available to their favourite app (for example their favourite CRM) but we also provide a portal to give access to sales and marketing teams to the insights they need (such as identifying the next best product to sell, whether their customer is at risk of churn, or if a portfolio monitoring is raised).
Conclusion
Helping financial services firms, life insurance companies and wealth managers scale their AI initiatives is at the heart of what we do. As a person willing to adopt AI to achieve more ambitious business objectives, you know the challenges that are in front of you and that many other FSI firms are facing. Leveraging Auto-ml can help you scale your initiatives by upskilling your business analysts rather than recruiting new expensive and hard to attract/retain individuals and to accelerate the work of your data-scientists while maximizing their chance to find an algorithm to maximize the ROI and success of your initiatives. Through our collaboration features you can easily set a diverse team of technical and business individuals and provide them with the tools that they need to take the data as it is and turn it into high value predictive models that can be used by your end users while respecting compliance and security. The data and your initiatives are in compliance with existing regulations at all times and on all geographies thanks to our embedded obfuscation, our multi-tenant architecture or role management functionalities. Finally, you can’t avoid setting an hybrid or multicloud strategy if you want to benefit from the flexibility and scalability it provides as well as modern chips (GPUs or chips optimized for inference) to power your more demanding workloads. Our solutions will help you support this strategy, providing you with the capabilities to build models in one cloud to deploy them in another cloud or on your own infrastructure.
