Dr. Hojjat Rakhshani

I spearheaded the development of an advanced AI-powered search engine to enhance the company’s e-commerce platform. This system leveraged state-of-the-art generative AI techniques to propose a semantic search engine capable of understanding user queries in context and delivering highly relevant product recommendations, significantly improving customer satisfaction and engagement.

We developed AI-driven solutions to optimize inventory assortment, directly contributing to €80 million in additional revenue while reducing stock costs in physical stores. This involved aligning inventory management with market trends and sales forecasts, enabling Decathlon to balance supply chain efficiency with customer demand effectively.

We built a comprehensive forecasting model using Amazon SageMaker DeepAR to predict store and product turnover accurately. This also involes analyzing the impact of Covid-19 on store performance using XGBoost regression, providing actionable insights that informed Decathlon’s strategic business decisions during a critical time.

In this project, we developed a pipeline to identify and classify relationships between scientific articles which achieved 90% accuracy, significantly enhancing the efficiency of literature reviews and research synthesis. The results were recognized in the IEEE WCCI 2020 proceedings, highlighting its potential to transform academic workflows.

This research focused on improving deep residual networks for time series forecasting through neural architecture search (NAS). This approach achieved state-of-the-art accuracy, surpassing leading benchmarks like HIVE-COTE. Published in the IJCNN 2020 proceedings, this work demonstrated how innovative neural architectures could provide more accurate and efficient time series analysis.

We investigated network interdiction problems in multi-depot vehicle routing, developing optimization strategies to improve logistical efficiency even under disruptions. This research combined advanced mathematical modeling and combinatorial optimization techniques, offering practical solutions for real-world transportation and supply chain challenges.

We designed a novel optimization framework that leveraged transfer and ensemble learning to reduce computational costs significantly. By retaining and reapplying knowledge from one problem to another, this approach improved efficiency without sacrificing performance, providing a powerful tool for tackling diverse optimization challenges.

We enhanced the Two-Stream Inflated 3D (I3D) deep learning architecture to predict crowd dynamics using the Crowd-11 dataset. By applying metaheuristic optimization, I fine-tuned the model for improved accuracy and real-time performance, contributing valuable insights to the field of crowd behavior analysis.

We proposed a multi-objective optimization framework to automate machine learning model configuration. By balancing competing objectives such as accuracy, computational efficiency, and robustness, this framework streamlined the deployment of ML models tailored to specific applications, saving time and resources while maintaining performance.