Abdou-Elbadi Jaouhar

— Designed, trained, and deployed vision models for agronomic use cases (Python, PyTorch, Triton Inference Server, TensorRT, CUDA, C++), with CI/CD integration and production monitoring.

— Designed a SaaS platform for image and metadata aggregation/annotation (NextJS, Postgres, Terraform, CI/CD) with manual and assisted annotation (pre-label via pre-trained model), integrated moderation; integrated Segment Anything for interactive segmentation.

— Deployed optimized models on NVIDIA Jetson (AGX) embedded platforms: ONNX → TensorRT conversion, pruning/quantization (FP16/INT8), memory and latency optimization; developed ROS/ROS2 nodes for real-time integration into the robotics stack.

— Implemented a dataset versioning system (dataset registry, schema validation, quality control), ensuring traceability and reproducibility of training and deployed models. — Implemented operational monitoring of models in production (Prometheus, Grafana) and integrated active learning mechanisms.

— Conceived the architecture and defined conventions for internal libraries, then led and co-developed their implementation with the team (roadmap, code reviews, releases, packaging, CI, versioning), standardizing training, evaluation, and deployment and reducing onboarding and time-to-PoC.

— Designed and operated training/testing/deployment pipelines (Vertex AI, Airflow, MLflow), promotion between environments and secure rollback.

— Led ML technical discussions, contributed to roadmap prioritization and product/tech alignment; collaborated with agronomy, mathematics, and hardware departments.

— Recruited, supervised, and mentored developers/interns/apprentices; conducted technical interviews.

Multispectral Imaging • Time-series •

Computer Vision

— Prediction of downy mildew evolution from time series (weather conditions, farm history).

— Detection of downy mildew on vines using multispectral and RGB images (structure tensor on aligned images).