Ouassila H.

-Implementation of a semantic matching solution between heterogeneous documents based on vectorization using Transformer models (GPT), and managing orchestration via LangChain and RAG modules. Embeddings are indexed in Vertex AI and the modeling of extracted entities is done with Neo4j in the form of a directed graph.

-Initiation and training of employees to use accessible AI to optimize their daily tasks, particularly Generative AI through a program spread over several days.

-Design and deployment of a scalable cloud infrastructure on Google Cloud Platform, designed to efficiently process massive volumes of data in real-time. This architecture ensured performance, flexibility, and optimized cost, while dynamically adapting to load peaks and business needs. Response times of less than 10ms in 90% of cases.

-Implementation of a data governance plan in close collaboration with business units in an environmental context, to ensure the reliability of repositories, improve traceability of processing, and optimize critical workflows. By co-building a data catalog, a business glossary, and performance indicators, we were able to model processes, automate workflows, and integrate tools like Talend and Power BI to manage workflows. This approach reduced processing errors by 30%, accelerated reporting by 25%, and significantly improved collaboration between IT and business units, while strengthening compliance with GDPR and ISO 27001 standards.

- Development of a solution for the 3D Bin Packing problem based on a hybrid approach combining several optimization algorithms: simulated annealing, genetic algorithms, and tabu search. Simulated annealing allowed for efficient exploration of the solution space while avoiding local minima, while the other methods were used to refine configurations and accelerate convergence. The average container filling rate reached 92%.

- Design and development of an image recognition algorithm based on YOLOv5 to detect in real-time the presence of safety measures on people. The algorithm uses object detection to identify people and equipment, combined with a multi-target tracking module (Deep SORT) to follow individuals across multiple images. Distance is estimated from the coordinates of the bounding boxes and camera calibration parameters, allowing an alert to be triggered if the safety threshold is crossed. The model achieved an average accuracy of 93% for PPE detection and 80% for distance estimation.

- Development of a budget optimization algorithm for a digital content management and distribution platform for businesses. Based on Marketing Mix Modeling (MMM), the model identifies the most effective channels and proposes an optimal allocation of resources to maximize content impact while reducing costs, following a data-driven approach.

- Leading seminars, conferences, and workshops to raise awareness about AI challenges and cybersecurity threats, addressing topics such as algorithmic ethics, risks related to deep fakes, and attacks targeting intelligent systems.

- Building and managing a team of data scientists, ensuring their technical supervision, skill development, and project coordination.

-• Within the framework of an automated infrastructure monitoring project, I developed Deep Learning models for image recognition applied to pipeline video analysis. Using Convolutional Neural Network (CNN) architectures combined with U-Net segmentation models, the algorithm accurately detects visual anomalies such as cracks, leaks, collapses, or obstructions. Videos from onboard cameras were pre-processed and then analyzed to identify risk areas. The system also integrates post-processing techniques such as morphological analysis and spatial filtering to refine results. The models achieved an accuracy greater than 80% for crack detection and 88% for leaks, with a false negative rate below 6%, enabling targeted and rapid intervention by maintenance teams.

-• Design of advanced predictive machine learning models applied to multivariate time-series data to address industrial and environmental issues such as proactive alarm management, optimization of chemical reagent dosage, and early detection of pesticide presence. By combining sequential modeling approaches like recurrent neural networks (LSTM, GRU), hybrid CNN-LSTM models, and algorithms like (XGBoost, Random Forest, LightGBM) integrated into automated processing pipelines. Feature engineering techniques such as sliding windowing, extraction of seasonal trends, and temporal normalization were used to improve prediction quality. Some of these solutions were deployed via containerized microservices, interfaced with interactive dashboards (Power BI), enabling agile decision-making and a tangible improvement in operational ROI.