Boehringer Ingelheim & E.ON SE.

Diabetic retinopathy screening needs specialists the world doesn't have enough of. Deep learning can classify retinal images at scale, but in a clinical setting a black-box prediction is unusable: a model nobody can interpret is a model nobody will use.

Started with a SHAP-based interpretability framework for clinical decision support models spanning linear, tree-based and deep architectures, with proof-of-concepts comparing XAI methods against established feature selection and subgroup detection approaches for clinical trials and drug development. My master thesis took it further: deep learning models, including fine-tuned ResNet50, for detecting diabetic retinopathy stages from retinal fundus images, with local colour enhancement and strategic data balancing driving the performance gains. The core of the work was trust: DeepSHAP explanations validated against clinically annotated lesions, so the model's attention could be checked against what ophthalmologists actually look for.

Financial performance across multiple products and countries, with data from 5+ customer solutions serving millions of users, and business decisions waiting on reporting that was slow to produce.

Led data acquisition and management across the customer solutions, streamlined the analytics pipeline, and published financial performance reporting comparing actuals against previous forecasts across products and countries. Designed new performance indicators that fed directly into business decisions.