An AI that lives in your WhatsApp and learns how you actually are in a relationship.
The only designer on the team. I work with the CEO on what the product is, and I own the front end on my own.
EWOR is Europe’s most selective founder fellowship, run full-time by unicorn founders who have built companies worth over €12 billion, SumUp and Adjust among them. It takes about 35 people a year out of more than 35,000 applicants.
The product is a ladder with three rungs. Understand somebody. Find them a good match. Bring the two of them together in real life.
I have worked on it twice. Both times the job was the same. Work out which rung had broken, and rebuild that one.
The short version. The first time, the top rung was broken: women were signing up and staying away from the events. I rebuilt the site and the mechanic around one audience, and female sign-ups went up 25%. The second time the first rung turned out to be too soft to stand on, so the CEO and I went back down to build it properly. The product now asks you for a real conversation and writes you back.
1Soul is a relationship AI that lives in your WhatsApp. You talk to it the way you talk to everybody else, it learns who you are from that, and when it has read enough it brings you and somebody worth knowing into the same evening. Romantic first, and friendship and work after that.
Getting there takes three things, in this order, and each one asks the person for more than the one below it.
Giving this thing my real conversations is safe
This AI’s judgement is worth listening to
A 1Soul event, or a meeting we arrange ourselves, is somewhere I can trust
I am the only designer here. I work with the CEO on what the product is and how it should feel, and I own the front end on my own, from the positioning through to the conversion number at the end of it.
I stay with a decision all the way down. Positioning, the design that carries it, the build, then watching what people do with it and changing it.
Round one is about the handover from rung two to rung three. People were being matched, and then staying home.
At the early events the split ran nine men to every woman. The sign-ups existed and the room stayed empty, and the site was the reason. Research called it a heavy magic book. It promised that an AI would date on your behalf, which is the one claim a woman deciding whether to walk into a room of strangers will not take on trust.
Bumble's founder announced that AI would take over the swiping. The comments under that post were the most useful research I did here.
People hand AI their calendar and their inbox. Romance is where they most want the other person to be real, and that gap is the design problem.
Our own users said the same thing about the old site, in their own words.
The objection is specific. She does not say the technology fails. She says it was forced on her, and that it sterilises the one thing she came for.
So the split holds up at scale. Women rarely object to the technology in principle. They are quicker to read it as a risk placed on them, and in dating, where the stakes are personal safety, that gap gets wider. My read was two-sided. There is a real opportunity for AI here, and the same evidence is a warning about how to enter. The job is to leave people room to feel in charge, so the AI never reads as something being done to them.
Then I watched how people arrive. Women join a dating app because it makes them feel good. Men join because women are already there.
The order matters, and most products have it backwards. Pull in a lot of men and you get volume with no retention. 1Soul is for serious connection, so we build for women first. The men come on their own after that.
So I set the positioning before anything was drawn. 1Soul is built for the university woman who wants something serious, and the product has to prove that in the first interaction. She is a student, so she lives inside a small world of shared campuses and mutual friends, and she has more to lose socially if an introduction goes wrong. That shaped everything after it.
I looked at the dating apps people already use.
Most run at six or seven men to three or four women, and that gap shapes everything else. Men send a lot of messages and get nothing back. Women open the app buried in messages from strangers. It gets worse for both sides, and the ratio keeps getting worse with it.
The same Pew survey found 68% of women say AI is moving too quickly, against 58% of men. A separate study of around 3,000 people in the US and Canada found women rate AI as riskier, and traced it to higher general risk aversion and greater exposure to the jobs AI disrupts first.
Eight in-depth interviews with the target audience, London university students, current and lapsed app users, and heavy users of the apps they had left. I ran and wrote up every one of them. AI clustered the transcripts by problem type, which surfaced eight areas. The top four:
This went straight into the words on the page. Three lines, written to take the edge off a woman's guard about AI coming into her love life.
Be seriously matched, not just casually seen.
Cupid helps match; you decide to love.
Cupid works all that in the background, you stay in the foreground to love.
Each one puts Cupid behind the person. It finds and it narrows, and the person is still the one doing the loving. That is the line I did not want the product to cross in its own copy.
So I rebuilt the site around one audience, the university woman in London who wants something serious, and I designed the mechanic to match. I took the shape from Secret Santa. You know your match is in the room, you get one small hint, and you decide when to open it.
What Cupid says before it asks her for anything.
You are in a room with people you already trust. The setting does the reassuring before anything is opened.
Your name is already on something. That certainty is what lets you enjoy the waiting.
One small hint, and you choose when to open it. Nobody opens it for you.
A dating app hands over the whole profile at once: the photo, the name, the hobbies. Then it asks a woman to judge a stranger off that card, and to go and meet him in person. She gets all of the surface immediately, and none of the depth. Secret Santa suggested another way in. Set it somewhere she already feels safe. Give her a little. Let her stay curious, and let her work out for herself which one is her Santa. The decision stays hers.
Research backed it. Women hold back from putting their own information out, photos most of all, and they will not let an AI date on their behalf. Being handed a finished answer feels like being managed. They want more real information about the other person, then the choice. Men are far more relaxed about showing themselves. The same feature lands in two different places, and the design has to hold both.
Slowing the first interaction down only works because of what 1Soul is for. Someone who wants something serious will wait for one good introduction. Someone who wants to swipe goes elsewhere, and that is fine.
Everyone starts in the same place. A man or a woman talks to Cupid, Cupid works out who suits them, and both sides get a match. If they want to take it further, we place them at a real event, something like a matcha party. Everyone in the room is a London university student, so the setting already carries trust a stranger meeting never would.
The room runs like Secret Santa. Everyone knows their match is there and nobody is handed a name. The woman gets the richer clue. She knows more about him than he knows about her, and she is the one who can move first. So she can watch the room, start the conversation when she is ready, or not at all. He cannot open a queue and he cannot skip ahead of her.
I wrote the feature list and I ordered it. This covers the people who sign up online and want to come to the group events offline. Not everything could ship, so I ranked each feature against one question: does this make the first interaction feel safe for her?
The rebuilt colour system: sage, blush, plum and wine in place of the old mysterious near-black and gold, directly answering the "spellbook," "untrustworthy" and "too heavy" feedback from research.
The whole of that version, as it shipped.
It worked. Female sign-ups went up 25%, users more than tripled, and at one event the women outnumbered the men for the first time. The room stopped being nine to one.
I ran the growth work alongside it.
I wrote the video concepts and decided what each had to do. Some showed the mechanic, because "you get a clue first" needs showing. I also set what we would never post: nothing resembling a swipe queue, and nothing that promised a match.
The biggest rule came from watching the Bumble post fail. We kept the talk about AI to a minimum. Leading with the technology is what set people off there, so our posts lead with the people and the evening and let Cupid stay in the background.
I worked on early growth as well. I ran an on-the-ground content strategy, which meant street interviews at London universities, timed so the filming and the sign-ups happened on the same day. I also wrote the playbooks for Instagram and TikTok, so each platform had its own plan.
Round two is rung one. Understanding somebody deeply enough to be worth trusting, and building the trust that makes it possible.
Clarity said the funnel was healthy. 5,503 sessions over ninety days, 83% of the page travelled, conversion where it should be and users growing steadily. On paper there was no reason to stop.
The events said otherwise. People signed up and never turned into a room. At the start all we asked for on WhatsApp was a handful of questions, and a dozen messages hold almost nothing of a person. So we knew them only in part, the matches we made were not good enough, and no couples came out of them. We were introducing people who barely knew each other, out of a pool far too small to choose from. The first rung was carrying all the weight above it and it could not take it.
Understand somebody
The second rebuild serves this oneFind them a match
Bring them together in real life
The first rebuild served this oneThe CEO and I agreed on the order. Build the first rung properly, get real depth on each person, then go after volume. Accurate matching comes out of those two and cannot be had before them.
Each version changed because of something we had got wrong, and the three homepages show it.
An AI that dates on your behalf.
Changed because nobody believed it, women least of all.
Changed again because a dozen messages hold almost nothing of a person.
The third one asks for something much bigger and gives something much bigger back. You hand over a real conversation, Soul reads all of it, and it writes you back.
I built it in the palette we already had, because keeping the brand consistent was the argument. On its own terms it worked.
Testing said no. I ran it with five people, putting our page next to an AI product with a clean, plain style and asking them to choose. The colour was good at mystery and romance. It was wrong for the thing we were now asking them to do. They called the tone and the site too serious, and every one of them picked the cleaner style.
So I stopped holding the old palette exactly and grew the system out from it instead. Parchment, a letter, a report you open. It points at the last time people got to know each other slowly, by writing, and it lets something invasive arrive as something intimate and worth trusting.
It arrives as a link, so it opens on a walk home or at two in the morning and travels into a group chat on its own. Then it opens as paper.
I never ask anybody for their data. I ask whether they want to know what Soul thinks of their ex. The conversation arrives because somebody is curious.
The first version moved on what people said in interviews. From the first rebuild onwards I had Clarity running on the site, so the second version moved on what people did, and that turned out to be the sharper instrument.
The most-clicked thing on the homepage was the arrow to the next couple. 49 clicks, 7.79%, against 20 on the main call to action. People were hunting for more proof that real human beings had done this, so I built them a wall of it.
Dead clicks sat at 10.87% across 598 sessions, on the testimonial cards. People were clicking the photograph expecting it to open, so I gave it something to open.
People travelled 83% of the page in 1.4 active minutes, and 90.81% had never been before. The homepage carries the argument alone, to somebody moving fast.
Two samples. The click percentages come from the desktop click map; the session figures are the whole site over ninety days.
So I cut the copy back, gave the real cases the room the copy had been taking, and made that part of the page something you move through rather than read.
Every part of it is answering something the research or the data turned up, so here is each one and the reason it looks like that.
The whole of it, as it shipped.
She gets a clue, and the profile stays closedEnough to be curious with, and it leaves the rest for her to find.
"You decide what you do, always", written into the onboardingA product that will not say it in words does not mean it.
Raw uploads deleted within 24 hours, stated on the pageThe promise is worth as much as the sentence that carries it.
Both sides consent before anything is readSomebody else is in that conversation too.
Analysis happens in session. The raw conversation never reaches our database, and the upload is cleared within 24 hours of the report going out. What survives is the report and what Soul learned about you, which is the part you asked for.
Deletion is the default. Nobody has to go and find a setting. The 24-hour window exists so a failed report can be regenerated without asking somebody to upload their conversation twice.
Somebody else is in every conversation that gets uploaded, and they are not the one who pressed the button. So the flow asks the uploader to confirm they have permission before anything is read, and the site says so on the page itself.
Across both versions one rule held. The AI works in the background and the person stays in the foreground.
Two products, one rule. That is the part I would bring to anything else.
Round one had one job: get the women who signed up to walk into the room. It worked.
Round two has a harder one. Get somebody to hand over the conversation they would only show a close friend. Over 5,000 people are using it now.
That number is the measure of the trust work. Every one of them read what the product was asking for, decided it was safe, and went and found the export button.
The second proof arrived on its own. People started sending their reports on: to a friend, into a group chat, to the person the report was about. Somebody who shares theirs has trusted it twice, and the second time is the harder one to earn.
Sharing brings in people who arrive already warm. The product started recruiting for us.
The first rung is where I am now, building the part that has to hold before any of the rest works.
I run and write up the interviews myself. The clustering, the cross-checking and the building run through AI, and the judgement stays with me.
The closer the work sat to a person, the more of it I did by hand.
The AI moved fast on everything downstream of a decision. It never made the decision.
Want to try 1Soul?
Sign up here ↗