borghei/Claude-Skills / engineering/senior-computer-vision
senior-computer-vision
VerifiedComputer vision engineering for object detection, segmentation, and visual AI, covering CNN and Vision Transformer architectures and ONNX/TensorRT deployment. Use when building detection pipelines, training models, or optimizing inference.
Senior Computer Vision Engineer
Design end-to-end computer vision pipelines for object detection, instance/semantic segmentation, and production deployment. Generates training configurations for YOLO/Detectron2/MMDetection, optimizes models for ONNX/TensorRT/OpenVINO runtimes, and builds dataset preparation workflows with format conversion and augmentation.
Core Capabilities
- Detection pipeline design — requirements analysis, architecture selection (YOLO/RT-DETR/Faster R-CNN/DINO), dataset prep, training config, and metric evaluation.
- Model optimization & deployment — baseline benchmarking, ONNX export, INT8/FP16 quantization, and conversion to TensorRT/OpenVINO/CoreML/TFLite per target platform.
- Dataset engineering — audit, cleaning, format conversion (COCO/YOLO/VOC/CVAT/LabelMe), augmentation config, and stratified train/val/test splits.
- Architecture guidance — detection and segmentation architecture trade-offs plus CNN vs Vision Transformer selection.
- Production targets — FPS, mAP, latency P99, memory, and model-size budgets for real-time, high-accuracy, and edge deployments.
When to Use
- Building an object detection or segmentation system from scratch.
- Optimizing and deploying a trained model to GPU, edge, or mobile.
- Preparing, converting, or auditing a computer vision dataset.
- Choosing an architecture for a speed/accuracy/deployment trade-off.
Clarify First
Before generating training configs or pipelines, settle the inputs below. When one is missing or vague and would change the result, ask about it; when it wouldn't, use a sensible default and say which one you used.
- Task: detection / instance or semantic segmentation / classification (selects the architecture and
--task). - Dataset: location and format (COCO / YOLO / VOC) to analyze or convert (the input to
dataset_pipeline_builder.py). - Deployment target: GPU / edge / mobile (drives architecture choice and
inference_optimizer --target).
Ask about the two or three inputs that matter most for this request, not the whole list. If the user wants a result straight away, produce it and state your assumptions at the top.
Tools
All three are stdlib-only CLIs. inference_optimizer.py needs the model's runtime (torch for
.pt, onnx/onnxruntime and numpy for .onnx) to analyse or benchmark; without it only
format-based recommendations are produced and --benchmark exits 2.
| Tool | Purpose | Example (runs as written) |
|---|---|---|
vision_model_trainer.py | Analyse a COCO/YOLO dataset and generate a YOLO / Detectron2 / MMDetection training config (--framework, --arch, --epochs, --batch, -o, --analyze-only, --json) | python scripts/vision_model_trainer.py assets/sample_coco_valid --analyze-only --json |
inference_optimizer.py | Analyse, benchmark and recommend optimisations per target (--analyze, --benchmark, --recommend, --target gpu/cpu/edge/mobile/apple/intel, --json, -o) | python scripts/inference_optimizer.py model.onnx --analyze --recommend --target edge |
dataset_pipeline_builder.py | Subcommands analyze, convert, split, augment-config, validate | python scripts/dataset_pipeline_builder.py validate --input assets/sample_coco_flawed |
Exit codes: 0 ok; 1 validate found errors (broken image references, malformed boxes);
2 bad input (missing path, unknown architecture or format) or a requested benchmark that
could not run. assets/sample_coco_valid/ validates clean (exit 0); assets/sample_coco_flawed/
has an orphan annotation and a 3-value bbox (exit 1).
Default recommendation: YOLO26 (Ultralytics' current family; COCO mAP50-95 yolo26n 40.9, yolo26s 48.6, yolo26m 53.1 per docs.ultralytics.com/models/yolo26/, checked November 2026), with YOLO11 kept for existing pipelines. The trainer's table cites its source per entry; check the docs again before quoting figures, since families turn over yearly.
References
Load the reference that matches the task — keep this file lean and pull detail on demand:
- references/detection-workflows.md — quick-start commands and the three end-to-end workflows (detection pipeline, model optimization/deployment, dataset prep) plus the architecture selection guide. Read when executing a pipeline.
- references/commands-targets-and-troubleshooting.md — framework command catalogs (YOLO/Detectron2/MMDetection/optimization), performance targets, anti-patterns, troubleshooting table, and success criteria. Read while running training or deployment.
- references/tool-reference.md — full parameter, example, and output-format reference for the three scripts. Read when scripting the tools.
- references/computer_vision_architectures.md — CNN backbones (ResNet, EfficientNet, ConvNeXt), ViT variants (ViT, DeiT, Swin), detection heads, and FPN/BiFPN/PANet necks. Read when choosing or tuning architectures.
- references/object_detection_optimization.md — NMS variants, anchor optimization, loss design (focal, GIoU/CIoU/DIoU), training strategies, and detection augmentation. Read when improving detection accuracy.
- references/production_vision_systems.md — ONNX/TensorRT export, batch inference, edge deployment (Jetson, Intel NCS), Triton serving, and video pipelines. Read when deploying to production.
Scope & Limitations
This skill covers:
- End-to-end object detection and segmentation pipeline design (data preparation through production deployment)
- Training configuration generation for Ultralytics YOLO, Detectron2, and MMDetection frameworks
- Model optimization and export to ONNX, TensorRT, OpenVINO, and CoreML runtimes
- Dataset format conversion (COCO, YOLO, Pascal VOC, CVAT), splitting, validation, and augmentation configuration
This skill does NOT cover:
- Generative vision tasks (image generation, style transfer, super-resolution) -- see dedicated generative AI skills
- 3D reconstruction, SLAM, or point cloud processing beyond basic depth estimation
- Medical imaging regulatory compliance (DICOM, FDA 510(k)) -- see
ra-qm-team/compliance skills - Real-time video streaming infrastructure (RTSP, WebRTC, GStreamer pipeline design) -- see
senior-devopsfor infrastructure
Anti-Patterns
Benchmarking Without the Runtime
Mistake: Reporting latency or speedup numbers from inference_optimizer.py when torch or onnxruntime is not installed.
Why it happens: The recommendations print either way, so the run looks complete.
Instead: Treat exit 2 from --benchmark as "no measurement". Install the runtime, benchmark on the target hardware, and quote only measured numbers.
Quantizing Without an Accuracy Check
Mistake: Shipping an INT8 model because it is 2-4x faster, without re-measuring mAP. Why it happens: Speed is visible immediately; accuracy loss shows up later, on rare classes first. Instead: Re-run validation on the held-out split after every precision change and accept the model only inside the agreed mAP budget.
Splitting After Augmenting
Mistake: Augmenting the whole dataset, then splitting into train/val/test. Why it happens: It feels like one pipeline step. Instead: Split first (stratified), keep val/test untouched, and apply augmentation online during training. Otherwise near-duplicates leak into evaluation and inflate mAP.
Files
| File | Purpose |
|---|---|
scripts/vision_model_trainer.py | Dataset analysis and training-config generation |
scripts/inference_optimizer.py | Model analysis, benchmark and optimisation recommendations |
scripts/dataset_pipeline_builder.py | Dataset analyze / convert / split / augment-config / validate |
assets/sample_coco_valid/annotations.json | Sound COCO sample (validate exits 0) |
assets/sample_coco_flawed/annotations.json | Flawed COCO sample (validate exits 1) |
references/*.md | Workflows, command catalogs, architectures, optimisation and deployment detail |
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |