Zain Hazzouri

Master Thesis: Exploring U-Net-Based Architectures and Explainable AI Techniques for Speech-Music Classification 12/2024 | TU Berlin

TU Berlin

• Developed and evaluated U-Net-based architectures (U-Net, Attention U-Net, R2-U-Net, and R2-Attention-U-Net) for efficient speech-music classification tasks.

• Utilized MFCCs, LFCC and other transformations as input features to improve audio classification performance.

• Integrated Explainable AI (XAI) techniques such as Grad-CAM to interpret model decisions and provide insights into feature importance.

• Optimized model performance using PyTorch and conducted rigorous experimentation on large-scale audio datasets.

Chat AI for Pipeline Configuration 02/2024 | Shell, TU Berlin

• Developed an AI chatbot similar to ChatGPT for automating ETL pipeline configuration.

• Used Retrieval-Augmented Generation (RAG) methodology to provide accurate and contextual responses.

• Integrated OpenAI API and Langchain for conversational capabilities.

• Utilized ChromaDB to create a highly efficient content store for real-time data retrieval.

Multi-Scale Speech-Music Classification 02/2023 | TU Berlin

• Designed and implemented a deep learning model to classify speech and music efficiently.

• Extracted audio features using MFCCs (Mel-Frequency Cepstral Coefficients).

• Compared and evaluated four U-Net model variations:

• UNet, Attention UNet, R2-UNet, and R2-Attention-UNet.

• Improved classification accuracy through model optimization and hyperparameter tuning.

Twitter Data Analysis for Medical Trends 2022 | TU Berlin & Accenture

• Built a web scraper to collect tweets related to medical science topics over a 1-year period for trend identification and analysis.

• Applied sentiment analysis and TF-IDF (Term Frequency-Inverse Document Frequency) techniques to analyze and rank the most popular and relevant terms in medical discussions.

• Processed and cleaned large-scale datasets using Python, Pandas, and NLTK, ensuring data consistency and reliability.

• Visualized insights with Seaborn and Plotly, highlighting trends and providing actionable insights into emerging medical topics.

• Conducted data-driven evaluations to support decision-making processes in collaboration with TU Berlin and Accenture, combining academic research with industry relevance.

Data-to-Text Generation for Task-Oriented Dialog Systems 2022 | TU Berlin

• Fine-tuned advanced NLP models, including GPT-2, T5, and BERT, to generate human-like responses.

• Designed a system capable of generating accurate and context-aware textual outputs for task-oriented dialog applications.

• Evaluated model performance using metrics like BLEU, ROUGE, and perplexity scores.

Image Inpainting 2021 | TU Berlin

• Implemented an U-Net Convolutional Neural Network for reconstructing missing parts of images.

• The model efficiently filled white gaps within images by learning spatial structures and textures.

• Improved model performance through optimization of loss functions and training with high-resolution image datasets.