AI that does one job and shows its work.

Five systems we build and run, how we check what they produce, and the rules that keep them honest.

What we run

Antesala and ConciergeAgents that answer customers on WhatsApp and the phone, using only information you approve, and hand off to a person when it matters.
EKO-IA team of agents that designs, builds and audits a website from a brief or a current URL, and emails you the preview.
EKO-IISecurity checks of the kind a researcher runs, explained in plain language and ranked by what to fix first.
ZeroMore than 460 AI models in one app, with the right one picked for each task.
Benchmark judgesTwo AI models from different vendors scored 771 agency websites. The note covers what they got wrong.

How we check AI output

A model that grades its own work flatters itself. These are the checks we put around AI output.

Measure what can be measured

Speed, accessibility and structured data are scored by script, not by opinion.

Judge the rest with more than one model

Judgment calls are scored by one model and re-checked blind by a model from a different vendor.

Rank head to head

Side-by-side ranking is more honest than absolute scores, which drift upward.

Report the disagreement

When the judges disagree, we publish a band and the disagreement, not a single rank.

Rules for AI that speaks for you

Approved information only

An agent answers from what you approved. It does not invent answers.

A person is one step away

When a question matters, the conversation goes to someone on your team.

Honest labels

If a product is beta, it says beta. If a number cannot be checked, it is not on this site.

Documented and recoverable

Every engagement ships with runbooks, recovery procedures and credential escrow.

We used two AI models to judge 771 websites. What they got wrong.

Absolute scores flattered us, head-to-head ranking was more honest, and the two judges only half agreed. What we would do differently if you copy this method.

Silent failures: the bug shape that costs us the most

Loud failures get fixed in an hour. The expensive bugs are the ones that report success. Four patterns we have hit while running our own products, and the one rule that catches them.