Halimi A.

Developed a complete solution for the automated analysis of identity documents:

Implemented an OCR Tesseract pipeline in Go, optimized by computer vision preprocessing.

Integrated vision-language models (VLM) via Nanonets to assist OCR in identifying areas of interest.

Developed and industrialized Faster R-CNN face detection models with fine-tuning for reliable facial area localization.

Developed predictive models within the Prognos® program to anticipate the degradation of critical systems in the Air France-KLM fleet.

Collaborated with business experts to transform technical signals into actionable indicators.

On-premise industrialization of a predictive maintenance solution applied to the air conditioning system of the Boeing 777.

I developed a method to recognize similar images using a multimodal approach based on SimCLR. The objective was to identify altered or cropped historical images while preserving their visual and contextual similarities.

Dataset assembly and enrichment: I built a dataset by combining original images and their modified versions. To enrich the context, I used natural language processing (NLP) models to generate descriptive captions for each image, thus providing an additional layer of contextual information for similarity matching.

Contrastive learning with SimCLR: Using SimCLR, I trained the model to differentiate between pairs of similar and dissimilar images. By combining visual features and contextual data, I achieved a high level of accuracy in identifying corresponding images across different variations.