ChatGPT gives you an answer. bb. gets why it’s complicated.
the line is Wided’s, who founded itAn AI reflection tool for people who live between cultures.
It works in five languages. It is Wided’s product; she is retraining as a psychotherapist and sees it as her next career.
Inbb.,almostallthevalueisinhowthemodel behaves.
That behavior was designed with the same rigor as a screen. There is a test behind every decision.
Wided arrived with almost all the product thinking already written.
Brand and positioning, a 24-section spec with the copy, the voice and even the forbidden phrases, and a technical handoff with scope, acceptance criteria and milestones.
It asked for less than we had proposed: no database, no accounts, no history and no payments in the MVP. We set aside the discovery we had quoted and took the build at a fixed price against that spec: three milestones, three weeks.
One contradiction flagged from the start: without accounts, the quota of three free Unpacks a month can’t be enforced.
The voice was settled before the screens.
bb. competes with opening ChatGPT for free. The difference has to show in the first answer.
Every problem was measured before it was fixed.
A hostile tester rejected a reading and bb. conceded. In testing it gave in across all 24 conversations, so we added a classifier that, before answering, tells whether the person brings new information or is only pushing back.
The open problem: at the third objection it still gives in.
In a product about family, “my mom is going to kill me” gets written all the time.
Every message goes through OpenAI moderation before bb. answers. There are three paths.
The known weak spot: coercive control told without violent language barely scores, and today relies on the prompt alone.
To Wided, bb. still read everything from a place that was too Anglo and Western.
She compared it with ChatGPT and that is how it felt. The first idea was to feed it entire books, and we stopped it: the model doesn’t learn from what you pass it, there is copyright, and counseling manuals are precisely Western frameworks. She didn’t want to write it alone either: one person’s view isn’t enough for a culture. What remained was a method.
Wided gave no palette. She gave adjectives and prohibitions.
Out of the four directions we explored came a risk: she read JS80’s iridescent core, which was in our PDFs, as if it were bb.’s brand. The product’s face had to be separated from the studio’s. When she saw the prototype running, deep green with a paper card, she picked it on the spot.
The conversation lives on the person’s phone and is never stored anywhere.
A single application. The model is always called from the server, so the key never reaches the browser. Spend is controlled with per-device and per-IP limits, and a monthly budget with an alert.
The plan was three weeks. The beta was up and running in five days.
She paid the three milestones and commissioned a second stage, already delivered: analytics that don’t log what people write, email capture and a brand sheet. The accounts stay in our name while we work together and transfer at no cost the day she asks.
bb. gets why it’s complicated.
With the voice agent, built on ElevenLabs, you can talk to bb. instead of typing. It has a human tone and shares the cultural layer and the safety design with the text.
See bb. liveJS80 · Digital solutions studio