Yacine Mezali

Managed a large-scale annotation campaign for an image classification dataset. Evaluated the robustness of CNN models using adversarial attacks (FGSM, PGD, CW) achieving 60% accuracy post-attack. Optimized image classification model performance to 95% accuracy. Developed a predictive model for forecasting future investments in France (Business France) with MAE: ~80 to 120 projects/year, MAPE: 10%-15%, R²: 0.75-0.85. Anomaly detection for GSM R receiver use case. Designed a reinforcement learning-based power grid control system over 200 failure scenarios, achieving an effective resolution strategy in nearly 90% of cases. Developed NLP-based tools to improve customer database quality (Enedis), achieving 90% F1 score for customer document classification. Designed predictive models for electrical connection forecasting (Enedis use case): MAE: 12 average absolute connections per zone per month; RMSE: 18 connections, capturing stronger variations in high-activity zones; MAPE: 8.7%, an improvement of approximately 22% compared to the historically used baseline model. Stack: Python, Pytorch, pandas, SQL, DQN, statistical modeling, time series, Deep Reinforcement Learning, gradient boosting, random forest, DQN, actor-critic, reinforce.

computer vision, anomaly detection, image processing

Managed a large-scale annotation campaign for an image classification dataset. Evaluated the robustness of CNN models using adversarial attacks (FGSM, PGD, CW). Optimized image classification model performance. Developed predictive models for business use cases (Business France). Anomaly detection for GSM R production use case. Stack: Python, computer vision, Pytorch, anomaly detection, image processing

Managed a large-scale annotation campaign for an image classification dataset. Evaluated the robustness of CNN models using adversarial attacks (FGSM, PGD, CW) achieving 60% accuracy post-attack. Optimized image classification model performance to 95% accuracy. Used YOLO for license plate detection. Developed a predictive model for forecasting future investments in France (Business France) with MAE: ~80 to 120 projects/year, MAPE: 10%-15%, R²: 0.75-0.85. Anomaly detection for GSM R receiver use case. Designed a reinforcement learning-based power grid control system over 200 failure scenarios, achieving an effective resolution strategy in nearly 90% of cases. Developed NLP-based tools to improve customer database quality (Enedis), achieving 90% F1 score for customer document classification. Designed predictive models for electrical connection forecasting (Enedis use case): MAE: 12 average absolute connections per zone per month; RMSE: 18 connections, capturing stronger variations in high-activity zones; MAPE: 8.7%, an improvement of approximately 22% compared to the historically used baseline model. Stack: Python, Pytorch, pandas, SQL, DQN, statistical modeling, time series, Deep Reinforcement Learning, gradient boosting, random forest, DQN, actor-critic, reinforce.

computer vision, anomaly detection, image processing