How do you fix what you can't measure?

A product leader inside one of the world's largest semiconductor and AI computing companies had been making the same case for years: developer-facing documentation across the portfolio was failing the developers who depend on it. They had no evidence to prove it. Too small for the seven-figure consultancies, too low-priority for internal shared services, the work was attempted in the margins for more than eighteen months: never staffed, never finished, never killed.

The deeper problem was cultural. Speed was the point: idea on Monday, release on Friday, often with no PRD, no personas, no jobs-to-be-done. But products that ship half-baked get relaunched, and launching three times is not fast. Debt piled up behind every release: deprecated platforms, stale docs, developers left holding the bag. Teams knew they were getting dinged for it. Nothing forced it to surface.

And the asset at risk is the company's most valuable one: the reputation of its platform among developers. You cannot fix what you cannot measure.

Two kinds of debt were compounding against it. Product debt: the readiness work promised every release and delivered in almost none. Organizational debt: what teams knew about product health, trapped in nine disconnected systems and in people's heads. Neither is a technology problem, and neither surfaces on its own.

Reframe the problem

Bad docs were the visible symptom of something upstream. We kept documentation as the entry point, since it is measurable and the pain is undeniable, while pointing the platform at the cause: releases shipping without the product fundamentals in place. That shift gave the initiative a business case instead of a complaint.

Prototype in the room

On day one of the onsite, the client put their head down for ninety seconds, looked up, and said the whole thing had to change. The Release Readiness dashboard was prototyped that week, in the room, and demoed to TPMs and content owners before anyone had time to write a requirements doc.

Build the measure that didn't exist

Sense, make sense, act. Nine sources normalized through a single API into 82M data points, then graded by a net-new, deliberately conservative model: LLMs at the edges where judgment is required, deterministic logic in the middle where it isn't, and a coverage-aware rollup so no product is graded as though missing data were good data. Six dimensions per product, one portfolio health score, page-level insight for the authors who have to fix it.

Run our own factory in the open

We built this with an agent swarm and felt the failure modes firsthand: agents with thin context on product intent and architecture produce work that looks production-mature long before the foundation is. A printing press guarantees volume, and volume alone. That is exactly what the platform now measures for: quality that scales at the same rate as speed.