Shams Roumili

Automation of Market Abuse scenarios

Summary: Built rule-based models to detect complex Market Abuse scenarios and integrate them into a user interface.

-Replicate the logic of the models (ex : Front Running) using Python based on specification documents, focusing on code optimization (pandas).

-Use Oracle to request input data (products, accounts, proprietary trades, customer trades) to be fed to the model, and manage configuration and results tables (creation, partitioning).

-Perform Gap Analysis with existing solution and suggest improvements.

-Communicate with users (ie : Compliance officers) on the platform and integrate their feedbacks.

-Optimize the model parameters to reduce the number of Low Level alerts.

Alert Filtering for Anti Money Laundering

Summary: Built a classifier for Money Laundering Alerts using Python to reduce the workload of AML analysts.

-Analyzed risk coverage of the current rule-based solution, and benchmarked other AML approaches

-Extracted transactional data into .csv files via SQL requests

-Created transactional and client features using Pyspark, and implemented supervised Machine Learning methods (sklearn) to classify the alerts between Non-issue and Escalated ones. Achieved a 42% reduction of False positives while maintaining a 95% Recall

-Ensured transparency and interpretability of the solution via SHAP values, to comply with Financial regulation.

-Constructed an unsupervised anomaly score (PyOD) to facilitate the manual processing of escalated alerts

Détection d'obstacles pour les voitures autonomes. Client : Toyota.

-Recherche, conception et implémentation de méthodes de détections d'anomalies basées sur la reconstruction d'images. Evaluation de ces méthodes (mAP, F1 score, IoU, temps d'inférence).

-Evalutation et adaption de méthodes supervisées pour la detection d'obstacles (YOLOv3) et explication des résultats par heatmaps d'activation et étude ablative sur l'achitecture.

-Identification des cas de défaillance (luminosité, occlusions, déséquilibres de classes et de tailles d'objets) et élaboration de stratégies pour contrer ces limitations.

-Création d'un dataset augmenté d'obstacles de route: 20% d'augmentation de l'AP.

-Intégration de features supplémentaires (profondeur, saillance, superpixels).