Team members:
Antal Nusselder, Lead Data Scientist, with:
Justin Smid,
Berlinda Hermsen,
Maxime Houtekamer,
Bart Santing,
Simon Dreyer,
Christiaan Zwiers,
Mariska Laagland.
Country: Netherlands
Organization: OHRA
OHRA is one of the largest direct insurers in the Netherlands and has been providing customers with 'surprisingly positive' experiences since 1925. OHRA customers can easily arrange all their insurance matters themselves when it suits them: by telephone or directly and easily with the OHRA App or via the website ohra.nl.
Awards Categories:
- Best Acceleration Use Case
- Best MLOps Use Case
- Best Approach for Building Trust in AI
Business Challenge:
Prior to using Dataiku, the process of building models within the central OHRA Data team in collaboration with other teams (e.g., the Actuarial team) and handing it over to IT to be deployed was manual, time-consuming, and difficult. With growing enthusiasm and demand for data products and data-driven solutions from colleagues outside of the Data team, OHRA was challenged to address both their desire to accelerate and their obligation to ensure that their solutions are responsible towards their customers and regulators.
The challenge to speak the same language
The Data team members come from different backgrounds with varying skill levels in programming and data science best practices. They had previously been developing models individually, each in their own way. This regularly led to differences in definitions. When collaborating with other modeling teams and even external collaborators it is now also much easier to work together and share and reuse code and model components that were previously not easy to integrate to tell one cohesive story.
Bottlenecks in the handover to IT to productionalize
When it came time to hand over their models to IT to be deployed in production, IT would often have to put in additional effort to retrofit, test and oftentimes rebuild the models in their own systems. Even though the responsibility and expertise of testing and validation of data products rests within the Data team, there would be an additional testing effort to ensure that the version of the model in production gave the same results as the version of the model built by the Data team.
Minimizing risks ahead when using AI and ML
On top of this, governing data products built on users’ laptops was entirely manual. If OHRA wanted to increasingly leverage AI and ML technology, it would be crucial to do so in a way that ensured they could be compliant with emerging European AI regulations as well as to be transparent and accountable towards their customers.
Business Solution:
The solution to the business challenge was to develop an integrated ecosystem that builds robustness while accelerating the time to value of their data projects. Within a year of adoption, they built this solution using the Dataiku platform.
The solution is made up of several parts:
- One common language: Dataiku provided the ecosystem for OHRA to standardize their data and model components within feature stores. This has greatly accelerated time to value as it has created a common base for all teams to have consistent naming, standards, and definitions of their data which is now pre-integrated, standardized, and automated. With the feature store, the Data team enables other teams to leverage data points and model components that have already been pre-calculated and adhere to the standardized definitions.
- Reusable Dataiku features: Working in a mixed-level team, OHRA has already built 10+ Dataiku plug-ins, generalizing the code that they write so that it can be re-used by their colleagues more easily. This empowers team members to collaborate on projects regardless of skill level. They have also developed macros to standardize templates in accordance with their governance framework.
- Operationalize trust & control: With the ambition to increase the usage of AI and ML within their business, OHRA created a governance framework that could be implemented within Dataiku to operationalize trust and maintain control over their data and data projects. This way, governance is embedded and standardized over all data products via prescribed controls, checks, processes, and roles and responsibilities utilizing custom code, plugins and the possibilities of Dataiku. This makes it easy for team members to check if the data products are made responsibly as well as for internal and external stakeholders. It also makes it very clear to the data scientists what is expected from them in terms of controls and documentation at every point in the development process.
- Custom metrics & Dataiku solutions: OHRA incorporated custom metrics and Dataiku solutions like the Model Fairness Report plug-in as integral parts of their data governance. This ensures bias assessment is applied in every use case that impacts customers, without additional time spent manually writing code or running queries and makes assessments reproducible and comparable between projects.
- Ease of deployment: Using Dataiku’s prediction and code API’s, the Data team is now in control of deploying their own models, without having to handover code or model artifacts to an IT team. This saves a lot of time in discussions and testing. The IT team simply calls the API through a mutually agreed upon specification. If the model needs to be updated or changed later on, the IT team can simply keep calling the same API without much additional effort. The controls built into Dataiku, together with the embedded governance provides the IT organization with trust that the data product they are integrating is built responsibly and reliably.
Day-to-day Change:
OHRA customers also benefit. OHRA can now test, build and deploy solutions that help our customers much faster. For example, automations in the claims process makes the claims handling more efficient and helps customers get answers quicker. With governance now standardized and embedded in the development process, OHRA can ensure their customers are treated responsibly and fair.
Business Area: Communication/Strategy/Competitive Intelligence
Value Generated:
Faster time to value & scalability
- Reduced development time: Development time has been greatly reduced with OHRA’s integrated solution. As input data is now largely centralized and standardized, there is no longer a need to do input data reviews to ensure data joins were conducted correctly. In addition, using visual, reusable components – either through standard Dataiku functionality or custom coded plugins – makes iterative development much faster. Not having to look up and write out code speeds up the process and makes code checking much easier, because the code used has already been approved. This allows the Data team to develop, check and bring to production data products at least twice as fast as before. It also makes it much more fun! Using the visual interface and recipes eliminates a lot of tedious code work and gives the team more time to spend on what’s actually important: helping our business and customers.
- Speed to production: Now that OHRA has moved from manual scripts to an integrated pipeline solution with enhanced collaboration, what would have taken one year to deploy a model in the past, now can take as little as two months. OHRA has greatly reduced time aligning priorities across different teams to facilitate handoffs. With the APIs set up within Dataiku, OHRA has a better link between the AI teams and IT, reducing the need for retrofitting and testing when moving toward deployment. They also no longer need to build new APIs from scratch for every model built.
- Scalability: Without the move to Dataiku as an integrated platform solution, the Data team wouldn’t have been able to meet the growing demand for data products from the business. They are now able to adopt a fail-fast strategy of quickly testing and prototyping to arrive at efficient business solutions. Deployed models are also much more scaleable. Using APIs instead of batch processes provides answers or predictions where and when they are needed. Using Dataiku, the underlying architecture is also easily scaleable when demand increases.
- Minimizing risk: With the implementation of custom metrics, the usage of Dataiku’s Model Fairness report plug-in, and the implementation of their Governance Framework within Dataiku, OHRA is also able to apply stricter, standardized and more transparent governance with little additional time spent.
Value Brought by Dataiku:
- Dataiku Business Solution: Model Fairness Report Plug-In: This is now an integral part of OHRA’s Data Governance and can easily be plugged into data projects.
- Visual Recipes / Visual Plug-Ins: Teams can collaborate more effectively regardless of their skill level as they leverage visual recipes and visual plug-ins. It also largely eliminates the need for code checking for these steps.
- Flow: Handover between colleagues is easier as Dataiku’s visual flow made it simple to see what a colleague is working on and take over work when needed.
- Feature Store: Onboarding is simpler as new colleagues can start working almost immediately. The learning curve of finding data and understanding how it connects is greatly reduced. Previously, new team members had to spend at least a month familiarizing themselves with the multitude of tables in the data warehouse and learning feature engineering standards.
Value Type:
- Reduce cost
- Reduce risk
- Save time
- Increase trust
