AI Scientist / ML Engineer

Eligibility: Candidates must be nationals of an EU member state, NATO member state, or NATO-allied
partner nation. Required due to defense/dual-use classification of projects (EU Dual-Use Regulation
2021/821, ITAR).

Own LUXUAV's sovereign AI capability: conduct original ML research, train and optimize proprietary perception
models on classified aerial datasets, and build the end-to-end infrastructure that productionises them. This role is
split between applied research (novel model design, domain-specific learning methods, adversarial robustness) and
MLOps (training pipelines, model CI/CD, edge deployment automation). All training and inference must remain fully
on-premise — no third-party vision APIs or cloud inference are permitted.

‍

AI Research & Model Development
- Conduct original research into perception architectures tailored to aerial defense use cases: small object
detection at variable altitude, multi-class threat classification, camouflage/occlusion robustness, and EO/IR
domain adaptation.
- Design, train, and evaluate real-time detection models (YOLOv8+, RT-DETR, custom architectures) and
propose novel adaptations where COTS solutions fall short of operational requirements.

- Research and implement domain adaptation and transfer learning strategies to maximize model performance
on limited, classified aerial datasets without external data augmentation from third-party sources.
- Develop active learning and semi-supervised learning pipelines to minimize annotation cost on high-security
datasets.
- Investigate and prototype multi-task learning approaches combining detection, segmentation, and scene
understanding in a single inference pass for onboard efficiency.
- Develop and maintain multi-target tracking modules (ByteTrack, StrongSORT, or custom) as standalone,
testable components delivered to the CV Engineer for pipeline integration.
- Maintain awareness of state-of-the-art ML research (NeurIPS, CVPR, ICCV, ECCV, ICML) and evaluate
applicability to LUXUAV's operational context.

‍
Edge Model Optimisation & Deployment Packaging
- Optimize trained models for onboard GPU hardware (NVIDIA Jetson, Qualcomm RB5) via quantization
(INT8/FP16), structured pruning, and knowledge distillation.
- Export and validate optimized models through TensorRT, ONNX Runtime, and OpenVINO toolchains; produce
deployment packages with performance benchmarks (latency, throughput, memory footprint) for handoff to the
CV Engineer.
- Define and maintain model performance contracts (accuracy floors, latency ceilings) gating production
promotion.

‍
ML Infrastructure & MLOps
- Design and maintain GPU training pipelines (multi-GPU, distributed training) on LUXUAV's on-premise GPU
cluster.
- Own experiment tracking, model registry, and reproducibility tooling (MLflow, DVC, or equivalent); enforce
dataset and model versioning across all AI projects.
- Build and operate CI/CD pipelines for model releases: automated regression testing, benchmark gating,
staged rollout, and rollback procedures.
- Manage the on-premise multi-GPU server cluster: resource scheduling, utilization monitoring,
driver/framework version management, and capacity planning.

‍
Synthetic Data & Digital Twin Integration
- Design and operate synthetic data generation pipelines (scene rendering, domain randomization,
augmentation) to supplement classified real-world datasets, feeding LUXUAV's Digital Twin program.
- Maintain dataset quality pipelines: annotation validation, class balance monitoring, and distribution drift
detection between training and operational data.

‍
Security & Data Governance
- Enforce air-gapped data handling protocols for all classified training datasets: access control, lineage tracking,
and audit logging.
- Ensure all ML tooling, model artifacts, and training infrastructure comply with data classification policies and
export control requirements.

‍

Required Qualifications & Experience:
- Master's or PhD in Machine Learning, Computer Vision, Applied Mathematics, or related field — PhD strongly
preferred for the research component.

- 5+ years of hands-on experience in ML research and model development for real-world vision tasks
(detection, classification, segmentation, tracking).
- Demonstrated ability to go from research idea to working implementation: paper reproduction, ablation
studies, and production-grade code.
- Strong proficiency in PyTorch; familiarity with TensorFlow/Keras and ONNX.
- Proven experience with detection/segmentation frameworks (YOLOv8+, RT-DETR, Detectron2,
MMDetection).
- Practical expertise in model optimization for edge: quantization, pruning, distillation, TensorRT/ONNX Runtime
export.
- Experience with domain adaptation, transfer learning, and data-efficient learning techniques.
- Hands-on experience building MLOps pipelines: experiment tracking (MLflow, Neptune.ai, W&B), dataset
versioning (DVC), and model CI/CD.
- Proficiency in Python (NumPy, SciPy, scikit-learn, Pandas) and pipeline automation scripting.
- Experience managing or operating multi-GPU Linux training environments (CUDA, cuDNN, SLURM or
equivalent).
- English: Upper Intermediate or higher.
- Good communication skills and ability to cooperate with adjacent engineering teams.
- National from a NATO member country or one of the following NATO Indo-Pacific partners: Australia, Japan,
South Korea, New Zealand or Ukraine.
- Free criminal record

‍

Preferred Qualifications & Experience:
- Track record of publication at NeurIPS, CVPR, ICCV, ECCV, ICML, IROS, or equivalent venues; open-source
contributions to ML or CV frameworks.
- Experience with active learning, semi-supervised learning, or self-supervised pretraining on domain-specific
datasets.
- Familiarity with aerial imagery specifics: overhead perspective, small object detection, multi-scale challenges,
EO/IR domain adaptation.
- Experience with synthetic data generation and domain randomization (Blender, NVIDIA Isaac Sim, Unreal
Engine/AirSim, or equivalent).
- Experience with adversarial robustness evaluation and model hardening for safety-critical deployments.
- Experience with distributed training frameworks (PyTorch DDP, DeepSpeed, FSDP).
- Familiarity with on-premise MLOps stacks (self-hosted MLflow, MinIO, Prometheus/Grafana for GPU
monitoring).
- Familiarity with DO-178C and MISRA C/C++ as they apply to ML model validation in airborne systems.

‍