07 / Architecture brief

Cloud portability: AWS → Azure

Source workload
A sovereign consultation-analysis platform for a New Zealand qualitative-research firm, forked from the UK government's i.AI Consult pipeline.
AWS evidence
Bedrock and EC2 p3-class built and tuned in development and performance testing; the rest inherited from Consult's upstream AWS stack.
Azure status
Architecture mapping only, not a delivery or certification claim.
Attribution
i.AI Consult's upstream design, extended by me — never claimed as built from scratch.
2 direct
AWS services used in my development and performance work
5 extended
AWS service groups inherited upstream and extended
7 mapped
workload capabilities translated to Azure
No hands-on
Azure delivery claim

A capability spine: which of the seven AWS services below I built and tuned myself, which ones I inherited from an upstream pipeline and extended, and where each one lands if the target stack were Azure instead.

1. Where the AWS side comes from

Most of this AWS work sits inside one paid engagement: a sovereign consultation-analysis platform built for a New Zealand qualitative-research firm (see the Stoa case study), architecture evaluated and operated in development and performance testing rather than in production. The platform began as a fork of the UK government's i.AI Consult pipeline, so a large share of its AWS footprint — ECS Fargate, Batch, SQS and the identity, secrets and event-glue layers around them — is upstream architecture credited to that project's origin, not something I designed from scratch. Bedrock and the GPU self-hosting work are the two pieces I developed and tuned directly.

2. Seven capabilities, AWS to Azure

Every AWS service below carries where it actually came from; every Azure service is an architecture mapping only, not a hands-on delivery claim.

  • Managed LLM inference: Bedrock → Azure OpenAI + Foundry catalog

    Bedrock is my own development and performance work: Llama 70B structured outputs and a tuned retry ladder, evidenced in Exhibit 02.

  • Batch pipeline jobs: AWS Batch + SQS → Container Apps Jobs + Storage Queues

    An upstream foundation extended: i.AI Consult's queuing design, adapted for legislative-scale batches.

  • Application hosting: ECS Fargate → Container Apps

    Another upstream foundation extended, credited to i.AI Consult's original container-hosting design.

  • GPU self-hosting: EC2 p3-class → NC-series VMs

    My own development and performance work: Ollama on Terraform-provisioned GPU instances.

  • Secrets: Secrets Manager → Key Vault

    Upstream foundation extended, inherited with the rest of the AWS stack.

  • Identity: IAM + Cognito → Entra ID

    Upstream foundation extended from i.AI Consult's design, not something I built from zero.

  • Serverless event glue: Lambda + EventBridge → Functions + Event Grid

    Upstream foundation extended, carried over from the original pipeline.

CAPABILITY SPINE / 01

Seven capabilities, three provenance lanes

The same seven mappings as three lanes — AWS provenance on the left, the workload capability in the middle, the Azure equivalent on the right — read across for the shape of the translation. The right-hand lane sits past the visible edge of the figure: scroll it sideways to reach it.

flowchart LR
  accTitle: AWS experience mapped to an Azure target architecture
  accDescr: Seven workload capabilities connect AWS to Azure. Bedrock and EC2 p3-class were used in Bart's development and performance work; the other five AWS service groups are upstream foundations he extended. Every Azure service is an architecture mapping only, not a hands-on delivery claim.
  subgraph AWS["AWS · provenance stated per service"]
    A1["Bedrock · Bart development/performance work"]
    A2["AWS Batch + SQS · upstream foundation extended"]
    A3["ECS Fargate · upstream foundation extended"]
    A4["EC2 p3-class · Bart development/performance work"]
    A5["Secrets Manager · upstream foundation extended"]
    A6["IAM + Cognito · upstream foundation extended"]
    A7["Lambda + EventBridge · upstream foundation extended"]
  end
  subgraph CAP["Workload capability"]
    C1["Managed LLM inference"]
    C2["Batch pipeline jobs"]
    C3["Application hosting"]
    C4["GPU self-hosting"]
    C5["Secrets"]
    C6["Identity"]
    C7["Serverless event glue"]
  end
  subgraph AZ["Azure · architecture mapping only"]
    Z1["Azure OpenAI + Foundry catalog"]
    Z2["Container Apps Jobs + Storage Queues"]
    Z3["Container Apps"]
    Z4["NC-series VMs"]
    Z5["Key Vault"]
    Z6["Entra ID"]
    Z7["Functions + Event Grid"]
  end
  A1 --- C1 --- Z1
  A2 --- C2 --- Z2
  A3 --- C3 --- Z3
  A4 --- C4 --- Z4
  A5 --- C5 --- Z5
  A6 --- C6 --- Z6
  A7 --- C7 --- Z7
A provenance-labelled map, not a delivery claim — see What this is not below.

3. What this is not

I have not run Azure hands-on. The right lane above demonstrates architecture translation, not delivery history: for each AWS service, the nearest Azure equivalent a comparable workload would use — not a claim that any of it has been built, operated or certified there.

4. Evidence

The AWS side of this map traces to running code and a paid engagement, not just this page.