Artificial Intelligence & Data Analytics

Given the multifaceted regulatory requirements of the EU AI Act and the complex internal structures of the bank, a holistic, coordinated approach was required that involved both the European subsidiary and the global entity.

Together with the client, d-fine developed clear designs for AI governance processes and guidelines, adapted existing governance processes and internal controls, and created audit-proof documentation and audit artifacts to ensure traceability, transparency, and preparedness for inquiries from regulatory authorities. All relevant provisions of the EUAIAct were systematically recorded and translated into technical and implementation-oriented requirements.

The bank's deposit model is based on a traditional ARIMA time series model. Due to strong external factors (pandemic, armed conflict in Europe, interest rate reversal), the quality of forecasts has recently declined, making a revision necessary.

In collaboration with internal (bank) and external (d-fine) experts in the fields of deposit modeling and artificial intelligence, we are revising the model by tapping into new data sources, clustering deposit segments, and evaluating various traditional and modern forecasting approaches. The best results are achieved with a moderately complex AI approach, while other (often even more complex) AI approaches, such as neural networks, show little improvement or even a deterioration in forecast quality compared to traditional approaches.

The bank is seeking to better consolidate its various in-house AI initiatives. The goal is to establish a structure that allows synergies to be exploited and resources to be used efficiently, while also promoting individual initiative and innovation. d-fine was therefore asked to develop a strategic concept.

A series of workshops was first held to clarify the company's vision for AI. A second strand of work was the establishment of a system for identifying and evaluating/prioritizing potential AI use cases. Recommendations on the technological infrastructure and sourcing model were also developed. The central aspect was the development of organizational structures that take into account both innovation management and AI governance.

d-fine supported a major energy supplier in planning and implementing a cloud-based AI platform. This support included strategic analysis of the feasibility and value potential of data analytics and AI approaches for optimizing energy trading. In addition to technical integration into the enterprise architecture and connecting numerous data sources, we also took organizational, cultural, and competence-related aspects into account.

A wide range of AI use cases have already been developed and operationalized on the platform, including applications for price forecasting, supply and demand prediction, investment evaluation, and improved operational modeling of wind farms. The use of proven MLOps processes ensures that the models can be continuously retrained with new data and effectively deployed in production. The predictions of the AI models flow back into the central data repository of energy trading and are available in curated dashboards and for free analysis to support trading decisions. The functional and technical operation of the platform was handed over to internal teams and is now being successfully continued by the customer.

As part of the legislative process, various interest groups are asked to comment on proposed legislation. These comments must be reviewed by the relevant ministry, and the key demands must be collected and evaluated. To support the process of consolidating this feedback, d-fine was commissioned to implement an application that uses large language models (LLMs) to extract the key demands and present them to the user in a frontend together with the corresponding reasoning.

The solution developed by d-fine is based on Microsoft's Azure OpenAI Services and groups similar demands from different associations together. An interactive front end allows the corresponding text passages in the original document to be highlighted for individual demands.

As part of a general revision of its model for forecasting the probability of default (PD model) for one of its sub-portfolios, our customer also wants to integrate information from customer communications into its model (in addition to classic, structured risk factors). d-fine supports the customer in extracting various information from email history and other communication channels and derives corresponding risk factors.

A particular challenge is ensuring compliance with data protection regulations. For example, references to illnesses may be given in email correspondence as reasons for deferral requests. To avoid processing this sensitive information in the process, d-fine implements a pre-processing step that recognizes this information and masks it for further processing steps.

The machine production of various machine components requires a glazing step to create a protective layer. The machines used in this process produce a high reject rate and experience regular but unpredictable failures.

d-fine supported the customer in developing a customized model for anomaly detection. The goal of the implementation was to predict a failure based on the machine's sensor data and identify the cause of the failure.

The model is based on a GAN architecture and was adapted to the customer's data. A special feature of this model is its inherent explainability (xAI). This made it possible to directly see why an anomaly was classified as such.

In addition, d-fine supported the customer in integrating the model into the Databricks data platform used by the customer and organized workshops for knowledge transfer.