Bananas president

Restaurants want menus in several languages, and the naive solution, pipe everything through a translation API, carries a failure mode best explained by an anecdote from a classroom: a girl wrote an essay about el presidente Maduro, ran it through a translator, and handed in a text about the bananas president, because “maduro” is also a ripe plantain ▸ 9:34. Machine translation doesn’t fail loudly; it fails fluently, in sentences that look fine to whoever can’t read them, which is precisely the person who needed the translation.

A restaurant cannot afford that ▸ 10:33. Dish names are brand; a mistranslated one is a joke at the owner’s expense, printed on every table.

So Picky’s translation design splits into two tiers with different trust levels, and the split is the teaching. Tier one: the languages the restaurant actually cares about get manually edited fields, a language switch in the admin view where the owner sees the menu in English, in Spanish, and hand-corrects every name and description; the human-verified version is the source of truth ▸ 10:41. Tier two: for the long tail, the Arab tourist who walks in unannounced, an on-demand AI translation to any language, useful precisely because the alternative is nothing ▸ 11:25. That translation can have errors, and that’s exactly why the two tiers exist as separate, labeled things ▸ 11:41.

verificado donde duele, automático donde ayuda →

The general form applies to any AI-powered product: don’t choose between human quality and AI coverage, tier them. Human verification where errors cost reputation; AI where the choice is imperfect help or no help at all; and an honest boundary between the two, so the fluent-sounding mistake never wears the badge of the verified one…