What is hard
Evaluation in a non-deterministic environment
The same question returns different brands each time. We run 11 engines from real devices in four regions, daily, and treat the output as statistics.
Credit assignment across weeks
Median time from publishing to the first citation is 14.5 days. The purchase comes a month after that. Attribution has to survive the gap.
Agents that touch production
We serve pages inside our customers’ own sites. The guardrails matter more than the model, and they have to be cheaper than the work they check.
How we work
Measure before you fix
If a problem does not show up in a metric, it is a preference.
The real web is the test set
Ship to customers, read the result in AI search and in sales, then change the system.
Agents do the work
People decide what the agent is allowed to touch, and write that down as a rule.
Contracts over conventions
What must not break is frozen as a test. A change that breaks one tells you what to fix.
Open roles
Technical co-founder
Own how the system is built: measurement, the content engine, serving, the event pipeline, agent interfaces.
GTM Lead
Close deals from the first diagnostic call and write the playbook for growing an owned site.
Customer Success Manager
Sit between the agent and the customer, and decide what still needs a person.
Employment form and compensation are decided in conversation, by fit and contribution. Korean is required for the customer-facing roles.
Open application
Tell us what you want to build, the hard problem you want to own, and the work you are proudest of. The first conversation is a 30 minute call.