Production-Ready Blueprints
Reference architectures and foundation baselines for public and private cloud platforms, data infrastructure, and AI runtimes.
When building platform foundations on Google Cloud, AWS, and Red Hat OpenShift, I build on official reference architectures rather than maintaining custom, redundant baselines. This aligns setups directly with upstream security, compliance, and vendor standards.
On top of these landing zones, I implement architectural patterns designed for data pipelines, policy-as-code enforcement, AI platforms, and private or sovereign environments.
Core Architectural Blueprints
- Secure Multi-Cloud Platform Foundations: Deploying secure, multi-account landing zones on GCP and AWS using Terraform / OpenTofu and Terragrunt. I implement strict Policy-as-Code guardrails utilizing OPA Rego, Conftest, and Kyverno to enforce organizational standards and VPC Service Controls.
- Private and Hybrid Cloud on OpenShift: Sovereign and regulated environments on Red Hat OpenShift Platform Plus — Advanced Cluster Management for fleet policy, Advanced Cluster Security, Quay, and OpenShift Virtualization for VM consolidation — delivered as GitOps-driven Validated Patterns with Argo CD, so one repository provisions datacenter, edge, and cloud clusters alike.
- Local-First to Scale Lakehouses: An agile analytical platform blueprint. Engineered for local-first prototyping with Polars, DuckDB, and Apache Arrow, it scales seamlessly to enterprise data lakehouses like BigQuery, Snowflake, or Databricks using dbt-core and Apache Iceberg v3 table formats, with DuckLake as the single-node catalog.
- Production AI Platform (LLMOps): Model serving on vLLM / llm-d behind the Gateway API Inference Extension, on OpenShift AI or plain Kubernetes with DRA-scheduled GPUs; Model-as-a-Service governance, OpenTelemetry GenAI telemetry, evaluation gates before every rollout, and policy-following guardrails (Llama Guard 4, gpt-oss-safeguard) — portable across GCP, AWS, and private cloud.
- Deterministic Multi-Agent Orchestration: A blueprint for production-grade AI workflows. Featuring stateful, cyclic multi-agent choreography using LangGraph and Pydantic AI, it incorporates Model Context Protocol (MCP) for system tool access, A2A for delegation between agents, and Langfuse/OpenTelemetry for tracing.
Recommended Cloud Baselines
These are the primary official frameworks I recommend and leverage:
- Enterprise Foundations Blueprint: Google's Terraform implementation of a multi-environment organization, landing zones, and secure-by-default networking and IAM.
- Google Cloud Security Foundations Guide: The definitive, official blueprint for implementing comprehensive security, identity, and compliance controls across Google Cloud.
- GCP Cloud Foundation Fabric: Google's official, highly modular Terraform-based framework for rapid prototyping and production-ready organization landing zones.
- GCP Cloud Architecture Center: Google's official catalog of cross-industry reference architectures, design patterns, and best practices.
- GCP Enterprise Generative AI Blueprints: Google's official reference architectures, notebooks, and production patterns for enterprise RAG pipelines, agents, and Vertex AI workflows.
- Agent Starter Pack: Production-ready agent templates (ReAct, RAG, multi-agent) with Terraform, CI/CD, evaluation, and observability wired in, targeting Cloud Run and Agent Engine.
- Gemini Enterprise Agent Platform — Agent Engine: The managed agent runtime — sessions, memory, code execution, and evaluation tooling — behind the Vertex AI agent stack.
- AWS Landing Zone Accelerator (LZA): The official framework for secure, multi-account AWS environments, designed to satisfy strict regulatory and security compliance frameworks.
- AWS Security Reference Architecture (SRA): AWS's official guidance and Terraform reference blueprints for designing and deploying security services in a multi-account architecture.
- AWS Well-Architected Tool & Catalog: The definitive framework for evaluating cloud architectures against cost, security, reliability, and operational excellence standards.
- Well-Architected Generative AI Lens: Design principles and eight architecture scenarios for generative AI workloads, from model selection to multi-tenant serving.
- Well-Architected Agentic AI Lens: The companion lens for autonomous agents — tool authorization, memory, isolation, and cost control.
- Amazon Bedrock AgentCore Samples: Reference implementations for the managed agent runtime, gateway, memory, and identity services on Bedrock.
- AWS Generative AI Application Builder: AWS's official reference blueprint and infrastructure-as-code patterns for building secure, multi-tenant generative AI applications on Amazon Bedrock.
- Red Hat OpenShift Platform Plus: The private-cloud baseline — OpenShift with Advanced Cluster Management, Advanced Cluster Security, Quay, and Data Foundation — for fleets that span datacenter, edge, and public cloud.
- OpenShift Virtualization: Runs virtual machines and containers on one KubeVirt-based platform, the standard consolidation path for on-premises estates.
- Validated Pattern: Multicloud GitOps: Red Hat's tested, GitOps-driven reference for managing clusters, applications, and policies from one console across clouds.
- Validated Pattern: RAG-LLM GitOps: A production RAG stack on OpenShift AI — model serving, vector store, and application — provisioned entirely through GitOps.
- Red Hat OpenShift AI: The Kubernetes-native AI platform — model serving on vLLM, pipelines, registries, and GPU scheduling with dynamic resource allocation.
- llm-d: The open source, Kubernetes-native distributed inference stack — disaggregated prefill/decode, KV-cache-aware routing, and multi-tenant flow control — that OpenShift AI ships.
- Red Hat AI Inference Server: Hardened vLLM with LLM Compressor for any open-weight model on any accelerator, on OpenShift, RHEL, or third-party Kubernetes.
- RHEL AI and InstructLab: A bootable foundation-model platform with the open InstructLab workflow for tuning models on private data.
- NVIDIA AI Blueprints: Reference workflows built on NIM microservices — enterprise RAG, multimodal PDF extraction, digital humans — deployable on any Kubernetes.
- Kubernetes Gateway API Inference Extension: The Kubernetes-native inference gateway — model-name routing, LoRA adapter awareness, and endpoint health for shared model server pools.
- CNCF Cloud Native AI Whitepaper: The CNCF reference for running AI workloads on cloud native infrastructure — scheduling, storage, networking, and observability.
- OpenTelemetry GenAI Semantic Conventions: The shared vocabulary for LLM, agent, and MCP telemetry — spans, events, and metrics — still in development status, pinned per release.
- Model Context Protocol (MCP): Anthropic's open standard for secure, deterministic agent-to-tool and agent-to-system communications, establishing modern integration baselines.
- Agent2Agent Protocol (A2A): The Linux Foundation standard for agents discovering each other and delegating tasks across runtimes — the counterpart to MCP.
- CIS Benchmarks (Center for Internet Security): The global configuration standards for securing cloud platforms, Kubernetes clusters (GKE/EKS/OpenShift), databases, and operating systems.
- Cloud Security Alliance (CSA) Cloud Controls Matrix: The cybersecurity control framework for cloud computing, mapping to major industry standards (ISO 27001, SOC 2, NIST SP 800-53).
- Open Policy Agent (OPA): The CNCF open standard for policy-as-code and uniform governance across cloud platforms, Kubernetes, and IaC templates.
- OWASP Top 10 for LLM Applications: The 2026 list of LLM application risks, with its companion Top 10 for Agentic Applications — goal hijacking, tool misuse, memory poisoning.
- NIST AI Risk Management Framework: The Govern–Map–Measure–Manage framework for AI risk, the reference ISO/IEC 42001 management systems and the EU AI Act crosswalk to.
- MITRE ATLAS: The adversary tactics and techniques knowledge base for AI systems, the threat-modeling counterpart to ATT&CK.
- MCP Security Best Practices: The protocol's own guidance on token passthrough, confused-deputy attacks, session hijacking, and authorization for MCP servers.