in buenos aires i learned to sell without credentials.
in silicon valley i learned that technology doesn't fix people.
in milan i learned that luxury protects what can't be copied.
in tuscany i'm learning not to confuse noise with progress.
STEWARD is live at https://t.co/dXA5R1egvo
It answers your customers from your docs, acts through your API with their own permissions, cites every source, and keeps a record of every turn.
One script tag to embed.
You pay only for delivered outcomes.
@xnavkum This is the question EU buyers bring to every security review now. Naming the rule behind a block is a good pattern. Answers need the same thing: the source and a record behind each one, so people can check it.
@the_jobman Very few. Buyers now ask the important questions: what does it read, where does the data go, and can you show me what it answered last Tuesday. The honest demo answers those first. Bonus track: now everyone think they can vibecode their "own AI"
@patcrowebar Good problem to pick. My bar for ready: real user questions pass, every answer points at its source, and someone can open the record a month later when a customer disputes it.
@techstarterio Trust will follow the proof. If the answer shows where the price and the slot came from, people will book. If not, they double check, and the shortcut is gone.
@GizmodoES el detalle que me preocupa no es el cheque. es que latam siga alquilando capacidad en vez de construir capa propia: datos, permisos, trazas. usar ia sin infraestructura es alquilar el futuro.
@swyx@thinkymachines the underrated part isn't cheaper tokens. it's that "same evals" quietly moves the bottleneck from model quality to judgment: who decides what good enough means inside the company.
@amasad chaos is the adoption signal. the companies that win aren't the ones that ban agents. they're the ones that give them a habitat: permissions, receipts, and someone who still owns the outcome
@emollick true. the scary version isn't deepfakes. it's when campaigns start optimizing for what the model rewards, and culture slowly starts sounding like a prompt that tested well.
@paulg the chart is the easy part. the hard part is the months before the first vertical spike, when the only proof you have is that you didn't quit the wrong idea.
@mehulmpt@maltekosub And the company on the receiving end still owns what it answered. When the customer is a machine, the record of what you told it and why matters even more, because nobody will screenshot it.
@eshamanideep Under 1% across 1.2M live turns is serious work. The last 1% is also where a source behind each answer pays off most, because you can find the bad turns instead of guessing.