Data Scientist — Computer Vision

Data Scientist — Computer Vision (Anomaly Detection in Drone Footage)

Stack: Python, PyTorch, OpenCV, Azure

The role

We’re building CV models to detect coating failures and external corrosion in industrial assets from drone imagery. You’ll lead experiments from dataset curation through model deployment on Azure.

What you’ll do

Own the CV experimentation loop: problem framing, baseline creation, hypothesis-driven iterations, and ablation studies.

Data pipeline: design ingestion, labeling specs, quality checks, and augmentation for aerial footage (motion blur, glare, altitude/scale changes).

Modeling: train/evaluate classification, detection, and segmentation models; compare supervised vs. weak/unsupervised anomaly detection when appropriate.

MLOps on Azure: reproducible training (Azure ML/Databricks), experiment tracking (MLflow/W&B), model registry

Evaluation & reporting: define metrics (PR curves, mAP/IoU, IoU/Dice, time-to-review saved) and build clear dashboards for stakeholders.

Collaboration: Work alongside domain experts and data scientists to reach modeling KPI’s. Partner with product/UX on review UI needs (markup, triage flows), and with domain SMEs to refine labels and acceptance criteria.

Performance & robustness: handle domain shift (sites, seasons, cameras), and optimize inference for high-res frames.

Required Qualifications:

End-to-end ML workflow ownership experience: data (exploration, augmentation), training, evaluation, deployment, with experiment tracking (MLflow or W&B).

Working knowledge of at least one of: image classification, object detection (e.g., YOLO/MMDetection), or segmentation (e.g., UNet/DeepLab/SegFormer)

Experience with Computer Vision evaluation metrics such as precision/recall, PR/mAP, IoU/Dice

Azure experience: Azure ML for training, Azure Blob/ADLS for data

Data ops & labeling: write clear label specs and drive quality (spot checks, IAA, active learning).

Strong collaboration mindset in a fast-paced environment

Clear, concise communication of experiments, trade-offs, and results.

Strong initiative to learn new methodologies and state-of-the-art modeling architectures

Preferred Qualifications

Domain familiarity with corrosion/coating inspection or industrial defect quality control.

Un/weakly supervised anomaly detection (autoencoders, student-teacher, one-class) and open-vocabulary grounding (CLIP, GroundingDINO, SAM).

Experience with CI/CD workflows in Azure

Practical DevOps and MLOps: Docker, reproducibility, versioning

Model / data drift monitoring and dashboarding experience

Inference optimization (ONNX/TensorRT), mixed precision, throughput/latency tuning.

Reviewer tools & human-in-the-loop workflows (FastAPI/Streamlit/Gradio).

Basic tracking for video/frame streams (e.g., ByteTrack/OCSort); simple multi-view or temporal fusion