AI roles in medical AI, satellite computer vision and banking: Cheikh Zaïd Foundation, GeoAP, Attijariwafa Bank and Arteka, with the measured outcomes.
Built an end-to-end medical-orientation RAG assistant on a two-stage hybrid retrieval architecture (FAISS + BM25 + cross-encoder reranking), served via FastAPI with rule-based symptom routing and emergency red-flag detection.
Designed an evaluation framework over a 1,000-question golden dataset with an independent LLM judge scoring faithfulness and hallucination, which surfaced a safety-routing bug and lifted emergency recall from 50% to 100%.
Hardened the assistant for production: prompt-injection filtering ahead of every LLM call, deduplication, version-safe index rebuilding, and a strict advisory-only workflow that defers to physicians and emergency services, covered by a pytest suite.
Built a two-stage chest X-ray triage pipeline (ResNet50 classifier + YOLOv8 detector) reaching 96.9% accuracy and 99.7% AUC on a 522-image held-out test set.
Designed a doctor-in-the-loop clinical workflow that persists physician corrections as structured data for monitoring and retraining.
Freelance AI Engineer — Satellite Computer Vision — GeoAP (Dec 2025 – May 2026)
Production MLOps pipeline detecting and segmenting agricultural and urban objects in satellite imagery, exported as GIS-ready GeoJSON.
Architected and delivered an end-to-end satellite computer-vision pipeline turning 180 source images into 12,000+ training tiles through overlap-aware tiling, with zero data leakage across the train/val/test splits.
Trained three complementary models over 10 object classes — YOLO11m detection (mAP50 0.836), YOLO11m instance segmentation (mAP50-mask 0.737) and a fine-tuned SAM1 ViT-H segmenter (mIoU 0.804) — covering buildings, trees, solar panels, swimming pools, cultivated parcels, rivers and roads.
Fine-tuned the SAM1 ViT-H mask decoder (4M trainable parameters, 300M-parameter encoder frozen) with pre-computed embedding caching across 9,958 tiles, cutting epoch time from 88 minutes to 5 on a T4 GPU, under a custom Dice + Focal loss with torch.amp mixed-precision training.
Exported inference results as GeoJSON FeatureCollections in real-world WGS84 coordinates, so detections and segmentation masks drop straight onto real maps in QGIS, Google Earth, Leaflet or any GIS-compatible tool.
Integrated SAHI (Slicing Aided Hyper Inference) tiled inference for robust small-object detection on high-resolution satellite imagery at arbitrary resolutions.
Shipped a FastAPI inference service with four endpoints and lazy singleton model loading, behind a dark-theme frontend with a live legend and GeoJSON download.
Configured a 12-stage DVC pipeline DAG with Roboflow dataset versioning, and GitHub Actions CI/CD running linting, 25 unit tests, API smoke tests and Docker build validation on every push.
Data Scientist — Attijariwafa Bank (Feb 2025 – Mar 2026)
Predictive financial models and internal AI tooling.