Saad Driouech

Developing diffusion-based generative models for GNSS signal reconstruction at Fraunhofer IIS, one of Germany's leading applied research institutes. - Iterating on model architectures and loss functions to improve training stability and reconstruction quality - Tracking experiments and monitoring training with TensorBoard - Working under real constraints: noisy data, low sample count

Built ML pipelines for garment return prediction in e-commerce, tackling a cold-start problem where historical user data is scarce. - 86% balanced accuracy on cold-start products - 5x preprocessing speedup after refactoring the pipeline - Feature importance analysis to surface actionable business insights

Contributed to PowerCARD, a global payment switching and card management platform. - Implemented compliance updates for VISA and Mastercard network requirements - Worked with SQL, CI/CD pipelines, Docker, and Linux environments

Research role focused on low-resource NLP for Moroccan Darija, a dialect spoken by ~30 million people with very limited existing language tools. - Pre-trained and fine-tuned transformer-based language models (52M–80M params) - 90% accuracy / 85% F1 on offensive language detection - Covered language modeling, text classification, and speech synthesis - Work resulted in 3 peer-reviewed publications

Taught Python and C/C++ programming labs to undergraduate students - Delivered hands-on lab sessions and evaluated student performance - Provided structured feedback to students and faculty - Co-designed a new instructional strategy to improve engagement and learning outcomes

Built a transformer-based text classification system to automatically routeclient inquiries to the right departments. - +7% accuracy and +4% F1 over traditional ML baselines - First end-to-end NLP project: from model development to integrated system