Azure Marketplace Readiness for Four AI Products

Four MII AI products packaged as Azure Managed Application plans and validated by real test installs. Releases are signed, ARM-TTK runs in CI and metering is built but switched off. Partner Center offers are not yet created.

36 defects
Install-blocking defects fixed before post-install checks passed
4 products
Managed Application packages validated by real test installs
1 day
Pivot from multi-tenant SaaS to Managed Application plans across four repositories

Role: Engineering lead for the packages, releases and test installs of four products. The offer model was my manager's decision, and the package layout follows a colleague's reference implementation.
Period: Oct 2026
Project type: Project
Room in the 3D HQ: L3:marketplace-productization

Results

36 defects
Install-blocking defects fixed before post-install checks passed
Across four products (14, 12, 5 and 5). Public and private plans were installed in a test subscription and torn down afterwards.
4 products
Managed Application packages validated by real test installs
Readiness only: Partner Center offers are not yet created, and metering has never called the live Metering API.
1 day
Pivot from multi-tenant SaaS to Managed Application plans across four repositories
12 stacked pull requests merged on 7 Oct 2026, the day the decision was made.

Problem

Metrodata wanted to sell four AI apps it built (voice contact center, code security, CV screening and fraud review) through Microsoft Marketplace, installed in each customer's own Azure subscription and billed by product usage rather than AI tokens, which customers already pay for directly.

Contribution

A multi-tenant SaaS track was built for all four products in one day. When my manager chose Azure Managed Application plans instead, I led the implementation of the pivot the same day: tenancy code deleted behind a fail-closed single-organization guard, with live data kept in place.
I built per-plan packages on a shared layout from a colleague's reference: an install wizard with region and Azure OpenAI quota checks, ARM-TTK in CI, and releases signed with non-exportable Key Vault keys through a GitHub OIDC signer without client secrets.
I implemented idempotent hourly metering in all four products: one record per dimension and hour, managed identity, product units only. It stays switched off until an offer exists.
I ran real Service Catalog test installs of public and private plans in a capped test subscription and fixed every install-blocking defect, including a managed resource group lock that stopped the apps' own identities from calling their AI services. For MII Fraud Detection, Starter public and Enterprise private test installs succeeded on Container Apps after five findings were fixed; its internal deployment still runs on App Service.