Mohamed Aymen Bouyahia

Permanent Job in the AI Team of the bank LCL (Credit Lyonnais)

– Multilabel classification of banking documents using the Donut model, trained on distributed GPUs with Accelerate library. – Use of the models gemini-2.5-pro, mistral-small and claude-3.5-sonnet-v2 for information extraction. – Working on GUI Agent using Computer-use repository made by Google Deepmind. – Development of an Agentic RAG to provide real-time, accurate answers for bank advisors' queries.

Detection of document fraud in medical insurance claims submitted by clients.

– Development of an OCR-based information extraction system using a voting mechanism integrating Tesseract, Doctr, and Donut. – Development of a ResNet-based autoencoder specifically designed to detect signs of forgery in documents. – Deployment of an angle deskewing tool (jDeskew) based on Fourier Transform. – Deployment of a perspective deskewing tool based on Segment Anything Model and openCV.

Development of an improved version of LIME, used in explainability, using Neural Decision Trees.

– Development of an interpretable deep learning model named Neural Decision Tree. – Elaboration of metrics to evaluate explanations: Fidelity & Stability. – Creation of library called LIME_NDT that uses Neural Decision Tree for explanation instead of Linear Regression.