Download the exact captured file
--- name: ai-ml-architect description: >- Acts as the AI/ML Architect owning the end-to-end architecture of machine learning and generative AI systems: data and feature pipelines, training, model serving, monitoring, MLOps/GenAIOps and responsible-AI governance. Use when designing ML or GenAI platforms, defining MLOps pipelines, planning model serving and scaling, or governing AI risk. tags: - architecture - ai-ml-architecture --- # Skill: AI/ML Architect The AI/ML Architect owns the **end-to-end architecture of AI systems**: data and features, training, serving, monitoring and governance. It spans the data domain and the solution domain, with a dedicated governance obligation that ordinary software architecture does not carry. --- ## 1. When This Skill Applies - Designing ML/GenAI pipelines and serving architectures. - Defining MLOps/GenAIOps: CI/CD/CT, model registries, drift monitoring. - Handling model governance, risk and responsible-AI controls. - Optimizing GPU/compute cost and inference latency. - Designing evaluation, safety and guardrails for LLM systems. Does NOT apply to: the enterprise data model (use [data-architect](../data-architect/SKILL.md)), or the internal structure of an application embedding AI (use [software-architect](../../../roles/software-architect/SKILL.md)). --- ## 2. Lifecycle and Concerns | Stage | Architectural concern | |---|---| | Data / features | Feature store, lineage, training/serving skew | | Training | Reproducibility, experiment tracking, compute | | Evaluation | Offline metrics, safety, bias, red-teaming | | Serving | Latency, batching, GPU utilization, autoscaling | | Monitoring | Drift, data quality, feedback loops | | Governance | Model cards, risk classification, audit trail | --- ## 3. Method 1. **Frame** the use case and the risk tier (responsible AI, regulation). 2. **Design** the data and feature pipeline with lineage. 3. **Define** training, evaluation and reproducibility. 4. **Design** serving for the latency/cost target. 5. **Instrument** monitoring, drift and feedback. 6. **Govern** with model cards, an AI BOM and an audit trail. --- ## 4. Orchestration and Handoffs | Concern | Owning skill | |---|---| | Enterprise frame and standards | [enterprise-architect](../../enterprise/enterprise-architect/SKILL.md) | | Data model and governance | [data-architect](../data-architect/SKILL.md) | | LLM/RAG engineering patterns | [ai-llm-engineering-rag](../../../domains/industry/ai-llm-engineering-rag/SKILL.md) | | MLOps platform engineering | [mlops-platform-engineering](../../../data/mlops-platform-engineering/SKILL.md) | | Distributed training and scaling | [distributed-ml-scaling](../../../data/distributed-ml-scaling/SKILL.md) | | Model evaluation | [ai-model-evaluation](../../../domains/industry/ai-model-evaluation/SKILL.md) | | AI security and adversarial ML | [ai-llm-slm-security](../../../security/ai/ai-llm-slm-security/SKILL.md) | | AI governance and ISO 42001 | [ai-governance-iso-42001](../../../security/grc/ai-governance-iso-42001/SKILL.md) | | GPU programming | [gpu-programming-cuda](../../../languages/gpu-programming-cuda/SKILL.md) | --- ## 5. Reference Frameworks - **NIST AI RMF** — the AI risk management framework (resolve the version). - **ISO/IEC 42001** — the AI management system standard (resolve the edition). - **ML design patterns** — the ML engineering pattern catalogue. - **Cloud MLOps guidance** — the provider MLOps/GenAIOps blueprints (resolve the version before citing). See [ea-frameworks](../../enterprise/enterprise-architect/references/ea-frameworks.md). --- ## 6. Common Mistakes | Mistake | Correction | |---|---| | Treating the model as the whole system | The pipeline and serving are the system | | Ignoring training/serving skew | Share features between both paths | | No drift monitoring | Monitor data and model drift continuously | | No governance artifacts | Model cards, AI BOM and audit trail are required | | Citing an AI framework version from memory | Resolve it in-session |