manisha.dev

Sanoma Media Finland, 2025

A second opinion for every AI decision

AI validation of advertising creatives, plus a Claude agent that reviews the AI's verdicts so human reviewers know when to look twice.

The problem

Every advertising creative that goes live on a large media network has to pass checks: does it match the booking, is the format right, is anything in it a problem. Reviewers in advertising operations did that by eye, at volume, under deadlines. Automating it with a vision model sounded simple. The hard part was trust. A model that is right most of the time still needs a person to know which verdicts to doubt.

What I built

I built AI validation of creatives with Azure AI Content Understanding, wired into the pipeline that carries a creative from booking to live delivery. Each creative gets analysed, its findings stored next to the booking, and the result surfaced in the AdOps review dashboard I built in Vue and TypeScript.

Then I built a prototype agent with the Claude Agent SDK that acts as a second reviewer. It reads the first model’s findings, the booking metadata and the creative itself, and returns an independent opinion with its reasoning. Where the two disagree, the creative is flagged for a person. Agreement lets reviewers move on; disagreement tells them exactly where to spend their attention.

How it is traced

Every step is a span. The validation call, the agent’s tool calls and its final verdict are recorded with OpenTelemetry, so a reviewer, or I at 3am, can replay why a creative was flagged. Retries are explicit, failures are visible, and the agent’s disagreements become a dataset for improving both models.

What it taught me

The model was never the bottleneck. The bottleneck was making its output something a busy person can trust or overrule in seconds. Designing the disagreement, not the agreement, was the work.