Content Accuracy Monitor | Frederico Luz - AI Engineer & Agent Architect

Facts expire. Articles don’t notice.

The blog runs on exactly those facts: mortgage rates, savings returns, tax brackets, application windows. And when the blog belongs to a bank, a stale number isn’t a typo. It’s a reader paying the wrong amount, or missing a deadline they thought they had.

CAM's dashboard, in Portuguese: 1,038 monitored articles; 75% of 6,279 facts confirmed current, 1,540 with an update ready and 54 still to confirm; 8,634 sources; 493 articles with updates, 26 with inconclusive facts and 0 new articles awaiting review.

I started by running AI validation on all the blog’s claims and finding the sources behind them. That gave CAM a baseline for each fact and the sources to watch for changes. Now the team can see exactly what its library promises, and how much of it still holds.

When something changes

CAM watches those sources, and the news, for anything that moves. When something does, every article that depends on it jumps to the top of the editors’ queue. Watching costs no tokens, so CAM runs thousands of source checks a day.

The editor checks the evidence and confirms; nothing reaches the blog until they do.

Interface real do CAM com dados fictícios: taxa de IRS de 21,5% para 21,2%, fonte oficial, trecho afetado e controlos editoriais.

Right number, wrong impression

A correct number can still mislead when its conditions are missing.

Interface real do CAM com dados fictícios: a dedução do IVA exige que o custo de aquisição do veículo, sem IVA, não exceda 50.000 €.

Only alerts worth reading

Early versions of CAM flagged anything checkable, down to Cristiano Ronaldo’s Instagram follower count, and every alert nobody needed made the next real one easier to ignore. So I set a stricter bar: if this fact were wrong, would it hurt a reader? If not, CAM stays quiet.

Interface real do CAM com dados fictícios: duas descrições do mesmo prazo e o controlo “Valores equivalentes” para dispensar o falso alarme.

When CAM still gets one wrong, the editor corrects it on the spot (same meaning, wrong reference, not worth tracking), and that mistake doesn’t come back. No tickets, no waiting on me.

What the team got back

There was plenty to find: CAM has confirmed newer values for 1,540 facts in 493 articles, nearly half the library. The editors now work through dozens of approvals and checks a day, and the repetitive part of the job, opening articles one by one and checking numbers against sources, is largely gone. What’s left is the part that needs an editor.

I continue to work directly with the editors to understand what they need and where CAM can improve. Their feedback shaped the product we have today, and it guides what we build next.