Lead Computer Vision Engineer

Designing, implementing, training, and optimising novel deep learning architectures for various computer vision tasks. You'll be comfortable with both framework internals.

Implementing complex image/video processing pipelines, optimising functions for performance (e.g., using CUDA modules), and handling advanced data augmentation.

Building custom containers, designing and managing complex training/inference pipelines (e.g., SageMaker Pipelines, Kubeflow), and managing cloud costs effectively for your team's workloads.

Setting up and managing shared experiment tracking servers for the team, creating standardised dashboards to compare runs, and enforcing best practices for experiment logging and reproducibility.

Designing multi-container applications using Docker Compose, deploying and managing services on Kubernetes using Helm charts, and ensuring efficient resource scheduling for ML workloads.

Designing detailed annotation schemas and instructions, writing scripts using platform APIs to automate quality checks, and developing strategies for active learning to reduce manual effort.

Core development for all computer vision tasks, data manipulation, scripting, and building custom tools and utilities for the team.

• Product Managers (for feature requirements and roadmaps)
• Software Engineering Leads (for integration into production systems)
• Data Science Leads (for data strategy and cross-functional research)
• Head of Engineering (for technical strategy and resource planning)
• Operations Team (for understanding real-world deployment challenges)

• Key Technology Vendors (e.g., cloud providers, hardware suppliers)
• Academic Research Partners (for staying ahead of the curve)
• Industry Consortia (for standardisation and collaboration)

More and more computer vision applications are moving to edge devices (e.g., smart cameras, drones, mobile phones) for privacy, latency, and bandwidth reasons. You'll need to become an expert in making models tiny, fast, and efficient without sacrificing too much accuracy.

Quantisation (Post-training & Quantisation-aware) · Model Pruning & Sparsity · Knowledge Distillation · Hardware-aware NAS (Neural Architecture Search) · TensorRT / OpenVINO / Core ML

• This week: Read up on the basics of model quantisation and its impact on performance and accuracy.
• This month: Experiment with quantising one of your existing models (e.g., using PyTorch's `torch.quantization` module).
• Month 2: Investigate a specific edge AI framework (e.g., TensorRT) and try to deploy a simple model to a compatible device.
• Month 3: Lead a discussion on the trade-offs between different edge AI optimisation techniques for our specific use cases.

Quick win: Take a pre-trained MobileNet or EfficientNet and try to run it on a Raspberry Pi or similar low-power device. It's a great way to get a feel for the challenges of edge deployment.

As our computer vision systems become more complex and critical, robust MLOps practices are essential. You'll move beyond basic model deployment to architecting continuous training, monitoring, and auto-retraining pipelines that ensure our models stay fresh and performant in production.

Data Versioning (DVC) · Model Registry & Versioning · Automated Model Retraining & Deployment · Model Monitoring (Data Drift, Concept Drift) · A/B Testing for Models

• This week: Familiarise yourself with tools like DVC for data versioning and MLflow's Model Registry.
• This month: Design and implement a basic continuous training pipeline for one of your team's models, triggered by new data.
• Month 2: Set up basic data drift monitoring for a production model, alerting you when input data characteristics change significantly.
• Month 3: Lead the implementation of A/B testing for models in a staging environment, ensuring robust evaluation before full deployment.

Quick win: Start versioning your datasets and model artifacts using DVC or MLflow for your next project. It's a small change with a huge impact on reproducibility.