Yexit · AI feature · 2025–now
Yexit keeps memories and messages for the people someone chooses. The Portrait feature collects small answers about a person so that, after their exit, Portrait AI can answer their family's questions in their words. I built the strategy: what a person is made of, how to ask without it feeling like a test, what to show back without a model, and how Portrait AI will be built and tested.
The people left behind ask concrete things: what she did on bad days, how she made big decisions, what she would say now. A personality quiz returns trait labels, which answer none of them. Two constraints made it harder.
The structure comes from personality psychology. Dan McAdams describes a person at three levels, in order of depth: traits, characteristic adaptations and a life story. Portrait asks about all three, and each of its twelve themes adapts a section of an established interview or inventory.
Twelve themes. A theme card opens on the Profile once it holds answers in two formats. People is an index across every theme, not a theme. Voice, the words someone keeps using and how long they write, is computed over all open answers. LSI: McAdams' Life Story Interview. AVP: the American Voices Project interview used by Park et al. IPIP: the public-domain International Personality Item Pool.
Every question is tagged with its theme, its level, its format and the interview section it adapts. The format is a second axis of the bank, never a substitute for the theme.
After each answer, one tap says how it felt: Easy, Mixed or Hard. It keeps the exchange closer to a conversation than a test, and it is stored with the answer.
Screens from the Portrait wireframes, draft v0.1. Sample answers throughout.
The flow keeps each step short. Counts and each theme's progress show from the first answer. A theme's card opens once the theme has answers in two formats, and the person's own taps on a card count as answers too.
Two rules protect the answers, and a third need pulls against them. Only one way of building the Profile met all three.
So I wrote the Profile's logic as a Python method. Every line is counted, not generated, and names the answers it comes from. An insight exists only where an answer has something to rub against: another answer, other people, or the person's own guess.
The trade-off is range. The Profile says only what can be counted, compared or matched to a written template. Anything that needs interpretation is left to Portrait AI, which exists only after release.
You decide rather than let it happen, 6 taps to 1. The exception was the move.
From 7 This or that tapsUntil then, the answers stay in the vault and nothing is generated from them. At release, Portrait AI follows the workflow of Park et al.: the person's own answers go into the model's prompt, and no model is trained on them.
Three versions of Portrait AI on the same model, each answering questions whose answers were held back, scored against the person's own answers. The score is divided by how consistently the person answers the same questions two weeks apart.
Planned for a first run on the user-test cohort, with one shared prompt wrapper, so that only the input differs between the arms.
1,052 people. Agents built from a two-hour interview, from survey answers (General Social Survey and Big Five), or from both. arXiv 2411.10109, v3.
Their three agents are the reason for the three arms: surveys for Profile, interviews for Stories, both for Combined. A written persona was the weakest input they tested, so the Profile arm gets raw answers. With 80% of a two-hour interview removed, their agents still reached 0.79. Yexit's small answers are far shorter than any of these inputs, so these results are a reason to run the arms, not a prediction of their scores.
The next evidence: whether people keep answering once the formats are short, how often a Profile line gets That's me, and what the three arms show.
Model names and plan figures are left out on purpose. The product is pre-launch.
Contact