LLM stands for Large Language Model, not Large Math Model. We learned that the hard way.

Hey everyone, I'm Benoit. I run a digital marketing agency in Quebec (Ursa Marketing) managing 200+ ad accounts across Google and Meta with a team of 20.

If you've ever run an agency, you know the tension. Every client expects their account to get the white-glove treatment. And they should. But 20 people across 200+ accounts means your team is spending most of their week pulling numbers, checking search term reports, and monitoring for the slow creep of ad fatigue, not building strategy or testing creative. The stuff they actually got into marketing to do.

So about a year ago, we started building AI agents to handle the monitoring and analysis layer. It did not go smoothly. Our first agent confidently reported a 8.4 return on ad spent (ROAS) on a campaign that absolutely did not have a 8.4 ROAS.

Turns out LLM stands for Large Language Model, not Large Math Model. 😅

We had to give a model that can pass medical exams a calculator just to divide two numbers.

So we rebuilt the whole thing. Deterministic code handles every calculation. Specialized agents each own a specific job (performance analysis, ad fatigue detection, search term audits). A second AI reviews every output before it goes anywhere. Nothing ships without passing that check.

A year in, every account gets consistent, thorough attention on a schedule. The team got 4 hours a week back per person to spend on the work that actually moves the needle for clients: strategy, creative, the stuff humans are good at.

We built it for ourselves because nobody else was building it. The platforms want you to spend more, not find waste. So we're turning it into a product and launching it here on Product Hunt next week.

If you're building in AI/MarTech, running accounts at scale, or curious about how AI agents can work in production without hallucinating your metrics, let's connect. More details dropping soon.