Content operations platform
Replaced three offshore workflows with a typed automation pipeline owned by an internal platform team.
- Fortune 500 retailer
- Automation
- 12 weeks
- Durable queues
- Typed contracts
- Design system
- Observability
Best fit when.
- 01Multiple vendor workflows need to be wrapped — not replaced — behind one typed surface.
- 02Your team owns the operational surface long-term but does not yet own a workflow-engine codebase to extend.
- 03Every stage requires an audit trail compatible with your retention policy.
What was happening.
Three offshore vendor workflows produced product copy, asset variants, and translated content for a large catalog. Each workflow had its own ticketing surface, its own SLAs, and no end-to-end view. Errors discovered late were expensive — by the time a translation regression surfaced, the asset was already in a marketing campaign.
Closing scope around this needed the discipline we describe in our writing on how we scope an AI engagement — vendor surfaces that 'might as well' have been included stayed out, on purpose, and the omission list was as long as the deliverable list.
What we were holding to.
- Existing vendor relationships could not be cut mid-engagement; the platform had to wrap them.
- The internal team had no prior workflow-engine surface to extend; the platform had to be greenfield, but small.
- Every stage needed an audit trail compatible with the brand's retention policy.
How we built it.
Typed contracts at every stage boundary
Every stage of the pipeline — intake, generation, review, publish — exchanged typed payloads with explicit invariants. Vendor adapters lived behind those contracts so a vendor swap is a code change to one adapter, not a workflow rewrite.
Durable, idempotent jobs
Stages ran as durable jobs with explicit retry and backoff policies. Idempotency was a first-class invariant — every job could be replayed safely, which made staging behaviour match production behaviour for the first time.
An operator UI built from the design system
Reviewers worked through a small operator UI built on the client's existing design system. We did not introduce new components or visual language — accessibility audits and brand reviews were already done.
What we left with the client.
- A deployed pipeline owned end-to-end by a named platform team.
- Typed IO contracts at every stage boundary, versioned and documented.
- Idempotency and retry policy written down, with a runbook covering the failure modes we saw in staging.
- Dashboards and alert routes wired to the platform team's existing on-call rotation.