A publishing control plane for 50,000-page releases
Sources, evidence, media, model work, editorial review, translation, scheduling, deployment, and measurement live in one operator workflow instead of a folder of scripts.
Validated 50,000-page streaming renderer · 50,000 conflict writes reduced from 6,558 ms to 1,010 ms median
How the pieces connect
- Source library
- Plan and generate
- Review evidence and media
- Schedule release
- Static publish
- Measure and correct
Rendered examples
The same source material in three working layouts.
The job
At scale, generating text is the easy part. The operation fails when source evidence, review state, media, cost, routes, deployments, and corrections cannot be inspected together.
The hard part
Keep human approval and deterministic checks in the path while still making large, multilingual releases practical.
How I built it
- Store evidence, generation inputs, outputs, costs, and review state as separate records.
- Make provider choice and model spending visible to the operator instead of burying them in background jobs.
- Render large releases as streams and validate routes, conflicts, links, and assets before activation.
- Keep publishing approval explicit; model output never becomes public merely because a job finished.
What I verified
- 50,000-page streaming renderer validation
- 50,000 conflict-write benchmark: 6,558 ms to 1,010 ms median
- Source, evidence, media, review, translation, schedule, and deployment surfaces
- Three neutral renderer families from one content fixture
Current state: The control plane and 50,000-page renderer are tested. The live portfolio cutover and growth data are still pending.