INETUM

Senior engineer with full, end-to-end technical ownership of the vision core of an industrial visual inspection product built on the NVIDIA platform. Combines deep expertise in unsupervised anomaly detection and defect segmentation with production-grade GPU inference optimization and complete model lifecycle management. Also able to design and lead the evolution of the pipeline orchestration toward a high-performance native C++ core, delivering real-time decisions on the factory floor.

- Expert-level PyTorch: CNN / transformer vision architectures, training, evaluation and rigorous ONNX export (zero train/serve skew).
- One-class / unsupervised anomaly detection: PatchCore, EfficientAD, PaDiM, student–teacher, normalizing flows, and their real failure modes (reference-set contamination, threshold calibration with few or no defective samples, synthetic defects via cut-paste / DRAEM, over-rejection, drift).
- Supervised defect segmentation and detection (encoder–decoder, DETR-family): training, acceptance criteria and imbalanced datasets.
- Methodological rigor: AUROC / AUPRO alongside plant-level metrics (escape rate, false-reject rate) and regression validation against recorded data.

C++ & Real-Time Systems

- Expert-level modern C++ (C++17/20) for real-time vision pipelines, in addition to expert Python.
- High-performance systems design: native pipeline/orchestrator coordinating capture, pre-processing, inference and post-processing while keeping data in memory and avoiding unnecessary copies and hops.
- Hard latency budgets:



determinism, watchdogs and graceful degradation.
- Linux, Docker, Git and CI as the natural working environment.

GPU & Inference Optimization (NVIDIA)

- Solid CUDA: execution model, streams, CUDA Graphs, memory management (pinned, unified, pre-allocation), writing and debugging custom kernels.
- GPU libraries: cuBLAS, NPP, CV-CUDA, Thrust or equivalents for accelerated image pre-processing and scoring.
- TensorRT in production: engine building, mixed FP16 / INT8 precision with custom quantization calibration, precision-degradation diagnosis and dynamic batching.
- Triton Inference Server in production.
- Profiling with Nsight Systems / Nsight Compute; p99 latency characterization per stage and finding the real bottleneck before optimizing.

Data & MLOps

- Model versioning and registry (MLflow or equivalent), reproducibility and dataset curation (CVAT).
- Traceability: able to demonstrate which model, data and version produced a given result.
- Models in production: monitoring, drift detection and a retraining / rollback policy.

Nice to Have

- Industrial cameras — GigE Vision (ideally Basler pylon); optics, lighting and photometric calibration (flat-field).
- Anomalib (advanced use or upstream contributions).
- Manufacturing / quality context (automotive or another regulated industry); ISA-95 and IEC 62443.
- Public cloud and cloud MLOps (ideally Azure: IoT Edge, ML).
- Publications, talks or open source in vision / anomaly detection.

#J-18808-Ljbffr

📌 Senior Computer Vision & Edge AI Engineer (Madrid)
🏢 Inetum
📍 Madrid