Dependable AI behaviour

Project description

As Product Lead at an AI company, I made persona behaviour something clients could trust, on their own platforms, not in a demo.

The complaints were never “the model.” They were vague: the persona sounds off, a safety check fires at the wrong time, a client wants something the product does not do yet. They needed it fixed. They needed to know it was fixed.

I turned every request into a ticket a developer could actually build: why it matters, what to ship, and the test that proves it is done. I tested the result on real conversations and blocked the release until it passed. Before every switch to a new AI model, I ran a test set for each safety check. When one client’s quality complaints piled up, I split them into specific causes, not a shrug and “it feels wrong.”

Then I wrote the guide and the changelog in plain words. No jargon. Short enough that a client knows what changed, who should use it, and who should wait.

The result: one client’s ask becomes a feature for every client. Nothing ships on a feeling.