Company Cooperation Project

Carrera
Inspired by data analytics in motorsports, a team of PwC data scientists and engineers set out to build a slot car racing track equipped with multiple sensors to showcase analytics capabilities at exhibitions. They developed innovative methods to control and monitor the cars while restructuring the existing backend into modular, domain-specific services. A central communication layer was introduced to manage data exchange between the frontend, backends, and the cloud. In addition, they designed an image recognition system capable of identifying the two race cars in a live video stream captured by a trackside camera, demonstrating advanced real-time data processing.

Project Anomaly - Outlier Detection in Utility Consumption Data
Reliable forecasts and robust monitoring are vital for energy distributors to manage peak demand effectively, ensuring grid stability and meeting customer needs, with a focus on detecting external factors influencing demand anomalies. Students developed a Machine Learning Pipeline to detect outliers and forecast utility consumption, showcased through a data-driven dashboard with essential KPIs for easy maintenance. Steps included data pre-processing, shock detection with peak categorization and explanations, model training and testing, and automated demand forecasting with dashboard integration.

Volatility & Volume
Students worked on a market trends project to understand market sentiment and investor activity across Xetra and Eurex. Project topics included clustering data, predictive power and analysing relationships between datasets. They used data science and machine learning to assess predictability and ultimately visualise the data with which a report was produced.

Intelligent Control on Efficient Anti-Icing(ICE-AI)
The main objective is to develop a computer vision model capable of detecting ice accumulation on the guide vanes of a gas turbine. During cold conditions, operators activate an anti-icing system that uses hot air from the combustor to heat incoming air and prevent blade damage. Since activation thresholds are set conservatively, this leads to efficiency losses. By integrating a vision-based ice detection model into the existing control logic, the system can be activated more precisely, improving turbine efficiency without increasing risk. Key activities include image processing, data augmentation, advanced modeling approaches such as 3D ResNet and anomaly detection, and experiments using infrared cameras to capture ice formation.

AI-Powered Cash Forecast
The project aimed to develop an AI-based model to predict customer payments and significantly enhance cash flow forecasting accuracy. A machine learning model was built using historical payment data to analyze and forecast customer payment behavior. The solution was integrated into the existing cloud infrastructure and tailored to company-specific requirements, including different organizational units and public holidays. Designed to be scalable and adaptable, the model can be continuously improved through retraining with new data. As a result, the company achieved more accurate cash flow forecasts, a flexible solution for various business units, and seamless integration into its existing IT landscape.

Digital asset return prediction
Navigating financial markets often involves processing vast amounts of information, making it challenging to take the right decision at the right time—especially in the highly dynamic asset class of digital assets. To support more informed investment decisions, Frankfurt School students enhanced the decision-making process by developing a forecasting algorithm to predict returns of individual digital assets using random forest models and neural networks. In addition, they built a sentiment indicator for major cryptocurrencies by analyzing Twitter data with natural language processing techniques, combining quantitative forecasting with market sentiment analysis.

NASA project on guided tranfer learning
In collaboration with NASA Ames Research Center and Frankfurt School, our project aimed to replicate and extend previous findings on using Guided Transfer Learning (GTL) to enhance AI performance on high-dimensional, low-sample-size (HDLSS) RNA-seq data. Given the challenges of overfitting in omics research, our approach builds on NASA’s work by pre-training AI models on large RNA-seq datasets (e.g., recount3) to capture gene expression patterns, enabling more efficient learning on small datasets. By refining GTL techniques, we seek to further reduce human intervention while improving few-shot learning, ultimately advancing AI applications in biomedical research.

Identifying Sanction Circumvention using Graph Analytics
In this project, we set out to revolutionize how project platforms inform decision-making for market positioning, training strategies, and project acquisition. Our team developed a scalable solution to extract, process, and analyze job/project postings, leveraging NLP techniques like tokenization, stemming, and keyword extraction. The result? Categorization of projects into focus groups (e.g., ML Engineering, Data Engineering) and a clear view of market trends over time.