cherif mohamed abderrazak

• Multimodal Data Alignment and Fusion – Designed and implemented a deep learning–driven pipeline for multimodal geospatial data alignment, combining Optical Flow, Stereo Photogrammetry, and non-convex optimization to achieve sub-pixel geometric consistency across modalities. – Developed hybrid C++/Python modules integrating PyTorch and custom geometry kernels (Boost, CGAL, Eigen) for efficient large-scale alignment and fusion. – Applied the framework to diverse domains including urban mapping (roads, buildings) and geoscience (fault detection), improving accuracy and scalability of automated map fusion. – Deployed the pipeline as scalable AWS microservices (AWS Batch, Lambda, Docker), orchestrated through Ray Serve for distributed inference and data streaming. • Linear Feature Segmentation and Vectorization – Led the development of a full deep learning workflow for line and fault segmentation from high-resolution optical and SAR imagery. – Trained and optimized Vision Transformer, SWIN, UNet, and Mask R-CNN models using self-supervision and domain adaptation to handle heterogeneous data sources. – Implemented an efficient C++ vectorization backend, transforming raster predictions into GIS-ready vector layers and reducing manual editing time by 60%. – Delivered containerized modules with automated inference pipelines and easy configuration for nontechnical users. • Software Infrastructure and Optimization – Built gRPC-based communication services for high-throughput processing of distributed geospatial datasets. – Refactored legacy pipelines into modular components with unit testing, CI/CD (GitHub Actions), and optimized runtime through CUDA and multi-threading. – Collaborated with researchers to integrate experimental deep learning architectures into production ready inference modules. • Tech Stack: PyTorch, PyTorch Lightning, Vision Transformers, UNet, Mask R-CNN, CUDA, C++17, Boost, CGAL, Eigen, Ray, gRPC, AWS (Batch, Lambda), Docker, FastAPI, GIS, OpenCV

• 3D Reconstruction from Satellite Imagery – Designed and implemented an end-to-end deep learning pipeline for 3D reconstruction from monoand multi-view satellite imagery, combining CNN-based disparity estimation with traditional stereo matching and photogrammetric optimization. – Integrated model inference with geometry post-processing (triangulation, meshing) to produce highquality structured 3D models from raw imagery. – Deployed the full 3D reconstruction workflow on MAXAR’s GBDX platform using Docker and AWS Lambda, enabling large-scale, event-driven processing of high-resolution datasets. – Collaborated with data and software teams to optimize I/O, caching, and tiling strategies, improving throughput and reducing manual supervision. • Change Detection and Vector Alignment – Trained and evaluated self-supervised and lightweight CNN architectures for optical image change detection, achieving over 90% accuracy across multiple geographies and sensors. – Implemented domain adaptation and continuous fine-tuning mechanisms to mitigate performance degradation under varying acquisition conditions. – Developed a geometry alignment module leveraging feature-based registration and non-rigid optimization to automatically align vector datasets, significantly reducing manual correction time. • Satellite Image Recommendation System – Built a hybrid recommendation engine to rank and select optimal mono/stereo image pairs based on geometric coverage, cloud content, and parallax quality using derivative-free optimization. – Developed an analyst-facing PyQt interface and integrated the recommendation core into a serverless AWS Lambda API for real-time automated selection. – Reduced analyst selection time from hours to minutes (semi-automatic mode) or seconds (fully automatic mode). • Tech Stack: PyTorch, PyTorch Lightning, OpenCV, Rasterio, QGIS, Geopandas, PyQt, AWS (Lambda, S3), Docker, CUDA, Stereo Matching, 3D Reconstruction, Self-supervised Learning, Domain Adaptation