Nexus

Tandem Interactive is a digital marketing agency in Fort Lauderdale, and Nexus is the platform the whole team works in. Client reporting and research and content production all happen inside it, with AI running underneath. I architected it and I write every line of it. Owning a system end to end is how I prefer to work, and on a platform this wide it also means nothing falls into the gap between two people's definitions of done.

Before Nexus everything lived where agency work always lives, which is a pile of browser tabs and exports that don't agree with each other, and now the whole team looks at the same live data instead. The AI works from each client's real numbers rather than estimates, and when data is missing it says so instead of filling the gap.

The rules underneath it

Client work runs on these whether anyone reads them or not, so they are worth stating plainly.

1. 01

   Every number carries its uncertainty.

   A confident point estimate is the kind of number that falls apart the moment somebody asks one good question. Nexus reports ranges instead, and only calls something a change once those ranges clear each other.

2. 02

   Nothing in the system is allowed to fabricate.

   A missing key or a dead data source produces a disabled feature with a label on it rather than a quietly invented number. A blank cell is information. A plausible wrong one is a liability that reaches a client before anybody catches it.

3. 03

   Anything with consequences leaves a trail.

   Actions that touch client data or spend money pass a permission check first and write an audit record after. The platform is fail-closed, so when it cannot verify something it refuses rather than guesses.

How the AI layer is put together

Model calls route through a layer that falls back when a provider degrades or a context window changes underneath it. Tool access runs through Model Context Protocol servers rather than bespoke glue written once per integration, which is what keeps adding a capability from turning into adding a maintenance burden. Anything with real consequences passes a permission check first and leaves an audit trail behind it, and the whole thing is fail-closed, so when it cannot verify something it refuses instead of guessing.

The AI visibility measurement inside it runs on the method I published in the writing here. Every prompt gets sampled several times per engine and every rate carries a Wilson interval, and a drop alert stays quiet until those intervals genuinely separate rather than firing on a single bad sample.

What I can share publicly

It's proprietary to Tandem, so this page stays in prose with no screenshots and no internals. What I will say is that building it alone means the data model and the AI layer and the interface were designed against each other rather than negotiated between teams, which is most of why it holds together.