About Me
Hi👋, my name is Rayane. I am a recent graduate holding an M.Sc. in Artificial Intelligence from ÉTS Montréal and an M.Eng. in Software Engineering from ISIS Castres (France).
I conducted my master's thesis at the Synchromedia Lab under the supervision of Prof. Mohamed Cheriet. My research lies at the intersection of Computer Vision and Multimodal Learning, with a specific focus on Document Understanding (DocVQA), where I worked on developing efficient and OCR-free architectures.
I am now open to new opportunities! My main interests include (but are not limited to) Visual Representation Learning, Model Optimization, and Explainability.
I conducted my master's thesis at the Synchromedia Lab under the supervision of Prof. Mohamed Cheriet. My research lies at the intersection of Computer Vision and Multimodal Learning, with a specific focus on Document Understanding (DocVQA), where I worked on developing efficient and OCR-free architectures.
I am now open to new opportunities! My main interests include (but are not limited to) Visual Representation Learning, Model Optimization, and Explainability.
Recent News
| 12/2025 | 🏅 Named to ÉTS Honour List (Fall 2025) Graduated with Excellent mention. Received Jury Recommendation for the Board of Director's Master's with Thesis Excellence Awards. |
| 11/2025 | 🎓 Master's Thesis Defense Successfully defended my thesis on OCR-free Document Understanding. |
| 10/2025 | 🏆 Best Paper Award at the VisionDocs workshop (ICCV 2025🌺) |
| 06/2025 | 🎉 Paper Accepted at ICCV 2025 Workshops (Spotlight) DIVE-Doc: Downscaling foundational Image Visual Encoder into hierarchical architecture for DocVQA has been accepted at the VisionDocs workshop. |
Selected Experiences
Graduate Researcher (Master’s Thesis)
Synchromedia Lab | Jan. 2024 - Nov. 2025
📄OCR-free Document Visual Question Answering (DocVQA) using Large Visual Language Model (LVLM) under the supervision of Prof. Mohamed Cheriet.
- Developed DIVE-Doc: A hierarchical end-to-end architecture reducing visual encoder parameters by 5x, halving the VE latency. Reached a 2.10 ANLS gap compared to larger teacher models (Paligemma) via knowledge distillation.
- Studied Fourier Feature for positional encoding: Integrated a spatially accurate positional module in the VE, improving performance from 82.67 to 83.46 ANLS.
- Adaptation to multipage document: Adapted a single page document model to process multipage without adding parameters, achieving competitive performance (71.73 ANLS).
Intern Data Scientist
Atout Majeur Concept | June 2023 - Aug. 2023
🏥Building a system to predict the Length of Stay (LOS) for incoming patients.
- Feature Engineering: Processed raw patient records and selected key clinical features to optimize model input.
- Model Training: Built an SVM model that uses patient symptoms and characteristics to make preddictions, achieving 78% accuracy with limited data.
- Automation: Developed an end-to-end pipeline to process new patient data and generate real-time predictions.
Intern Data Scientist
CHU Toulouse | May 2022 - Aug. 2022
🚑Developed a forecasting system to predict the volume of 911 emergency calls for the upcoming week based on historical data.
- Data Management: Engineered and preprocessed emergency call datasets from SAMU31 (french 911).
- Data Analysis: Performed exploratory data analysis (EDA) with geospatial and statistical visualizations, revealing why certain smaller cities experienced higher call volumes than larger metropolitan areas.
- Modeling: Designed and compared statistical (ARIMA) and deep learning (LSTM) architectures for time-series forecasting, reaching 80% prediction accuracy.