senior-computer-vision_skill

Overview

This skill delivers world-class computer vision expertise for building, optimizing, and operating production-grade image and video AI systems. It covers model development, real-time inference, 3D and video analysis, and end-to-end deployment using PyTorch, OpenCV, YOLO, SAM, diffusion models, and vision transformers. The focus is on practical, scalable solutions that meet latency, throughput, and reliability targets for production environments.

How this skill works

The skill inspects project requirements, designs appropriate vision architectures, and implements training, evaluation, and inference pipelines with production-grade tooling. It optimizes datasets, model architectures, and inference paths (quantization, batching, hardware offload) and integrates monitoring, CI/CD, and security practices. Hands-on scripts and patterns streamline dataset processing, model training, deployment to containerized clusters, and runtime monitoring.

When to use it

• Building or re-architecting a vision system for production
• Implementing object detection, segmentation, tracking, or 3D vision features
• Optimizing inference latency, throughput, or cost at scale
• Designing CI/CD, monitoring, and automated retraining for models
• Deploying real-time video pipelines or distributed vision workloads

Best practices

• Adopt test-driven development and version control for data, models, and code
• Design modular pipelines: dataset ingestion, augmentation, training, serving, monitoring
• Profile and optimize inference: quantization, batching, hardware placement, caching
• Automate deployment with containers, Kubernetes, and canary or blue/green releases
• Instrument models and data flows for drift detection, latency, and error monitoring
• Enforce security and privacy: auth, encryption, PII handling, and compliance checks

Example use cases

• Real-time object detection on edge devices with YOLO and quantized PyTorch models
• High-accuracy instance segmentation for manufacturing QA using SAM and custom finetuning
• Video analytics pipeline for multi-camera tracking and anomaly detection with distributed processing
• 3D reconstruction and pose estimation for robotics or AR applications
• Production deployment of a vision microservice with autoscaling, observability, and automated retraining