Hamza Boughanim

Architected and delivered a production-grade MLOps satellite computer vision pipeline for detection and semantic segmentation from high-resolution imagery. Processed 180 source images into 12,000+ training tiles using overlap-aware tiling with leak-free train/val/test splits. Trained complementary models (YOLO object detection, YOLO instance segmentation, and fine-tuned SAM1 ViT-H) across 10 classes, implemented SAM fine-tuning with embedding caching and mixed precision to reduce epoch time, and exported GIS-ready GeoJSON outputs in WGS84. Built SAHI tiled inference for robust small-object detection and served models via a FastAPI REST API with lazy model loading and a frontend legend with GeoJSON download. Set up a DVC pipeline (12 reproducible stages) with Roboflow versioning and GitHub Actions CI/CD including linting, unit tests, and Docker build validation.

Architected and deployed an AI-powered document automation platform integrating OCR, NLP, and vector search for insurance workflows. Built a hybrid OCR engine (Tesseract + EasyOCR) for Arabic/French extraction, created a Pgvector semantic search system with HNSW indexing for sub-second retrieval, and integrated Phi-3 via Ollama for local NER and document classification. Implemented secure S3 storage with AES-256 encryption, audit trails, and zero-trust principles. Delivered a web app for upload, semantic search, and validation with a chatbot assistant. Engineered automation for insurance notice generation and optimized SQL Server stored procedures to reduce production latency by ~30%.