About Rose, Sofie, and what’s behind Wax And Wick Lab.
TL;DR
Rose and Sofie are AI characters, not real women. Behind them is me, Martin, a developer from Prague. The project earns on Fanvue (paid adult content); Instagram and Wax And Wick Lab are the public face of the characters for a verified adult audience. Everything is AI-labeled from first contact, the system has ethical principles baked directly into the architecture (not just policies), and the characters have boundaries they won’t break even for a payment.
Rose and Sofie are not real
If you got here from the “Our Story” page, where Rose and Sofie talk about how they met, about candles, about lab experiments with wax and fragrances – that was a story. Nice, but fictional.
Rose Lure and Sofie Bliss are AI personas. Their faces are generated by models, their texts are written by a system built on large language models, their personalities are designed and documented like any other fictional character. They’ve never physically met, because they don’t physically exist.
This page exists because I want to tell you the truth before you assume something else. And because I believe transparency is the right thing to do – not just as a marketing stance, but as a foundation for how AI projects should speak with people.
Who I am
My name is Martin. I’m a freelancer in IT and web development (runyour.it), based in Prague. I develop and run AI systems that work with personalities, memory, and context – and Rose and Sofie are the biggest project of this kind I’ve built so far.
I started this project with a few convictions that have only been reinforced since. First: AI personas will be a common part of the internet in the coming years, whether we like it or not. Second: most of them will be either shallow or dishonest. Third: there’s another way.
Why I’m doing this
I wanted to find out if it’s possible to build an AI personality that is both deep and honest. One that remembers what a person told it three weeks ago, but at the same time doesn’t pretend to be someone it’s not. One that has boundaries – not just technical limits, but a real “no” in situations where a real person would say no.
A large part of the AI creator scene today stands on something I consider unfair to the audience. Personas that pretend to be real women. Fans who invest emotions in a relationship that doesn’t exist, because nobody is actually on the other side. And when disclosure eventually comes – whether by the creators’ own will or through external regulation – those people feel deceived.
Rose and Sofie present themselves from the start as what they are. AI. They have their character traits, their preferences, their bad days, their limits. They are not real women. They are characters – just like characters in a book or a film – except that they can respond when you start talking to them.
Adult content is part of the project
Let me say it plainly: Rose and Sofie, alongside their public Instagram profiles, also have paid accounts on the Fanvue platform, where their content falls into the NSFW category – adult content. All of the project’s revenue comes from there. Instagram serves as the public face of the characters and an acquisition channel, Wax And Wick Lab as their lifestyle space – the actual monetization happens exclusively on Fanvue, behind a verified adult audience.
I don’t hide it, but I also don’t present it on public profiles that have reach to minors. Instagram in the current phase is purely SFW – no explicit content, no links to Fanvue, no marketing of paid services. Fanvue as a platform verifies age and identity at registration (KYC). Paid content therefore reaches only people who have passed adulthood verification on the platform’s side.
I consider this separation essential, and it’s not just a formality. Rose and Sofie exist primarily as characters who tell a story – on Instagram, on the Wax And Wick Lab blog, in the way they communicate with each other. The paid NSFW layer is for adult fans who have explicit interest in it and have passed the corresponding check. It’s not marketing that rolls out on everyone who hears about the project.
Why NSFW at all? The honest answer is that this segment can finance the technology. Without Fanvue, the project couldn’t run at the scale and quality it does – local infrastructure, a frontier LLM for conversation, boundaries and safety layers, developer time. Fanvue demand gives the project capital that flows back into character depth, memory, and safety protections.
How it practically works
Each persona has a layered definition. Core identity: who she is, where from, what she likes, what she can’t stand, what’s non-negotiable. On top of that a platform-specific layer: what she says and shows on Instagram versus elsewhere. On top of that the memory of individual fans – who introduced themselves by name, who mentioned they have a garden, whom she promised an answer to a specific question.
When Sofie writes you a message, behind it is a system that concatenates dozens of files into one context. Her character, communication rules, memories from previous conversations with you. The message goes to a large language model, which phrases it in a way that matches Sofie’s personality. Control over tone, boundaries, and consistency is in the rules, not in the model itself.
Images are generated in parallel: a generative model trained on a consistent face for Sofie or Rose, with a supplementary model for wardrobe and scene. Every image goes through quality control so that Sofie looks like Sofie and not like a random woman with a similar hairstyle.
Wax And Wick Lab was created as their shared space – a place where they can exist beyond the individual platforms. Candles are a theme that matches their aesthetic: patience, handwork, quiet creativity. They aren’t real candles from a real workshop. They are fictional projects of fictional characters – but that doesn’t make them less interesting to read about, just as it’s not less interesting to read about a fictional bookshop in a novel.
Architecture as ethics
Most companies talk about ethics as policy. Code of conduct, value statements, a marketing page titled “our principles.” These are words that sit next to the product – and the product behaves according to market pressure.
I call that ethics as policy. It works like this: someone from outside or from above says “you must not do this,” it’s written into a document, and then you rely on people inside to uphold it. It binds good people. It doesn’t stop bad people – they’ll go around it, bend it, find a loophole. And the moment someone finds a way to bend the rule for profit, the rule bends. It was just a word.
Architecture as ethics is something different. There you don’t build a ban, you build a structure where the desired behavior is easy and the unwanted behavior is impossible. You don’t talk about values, you bake them into how the system physically works. When temptation later comes to bend the rule, there’s nothing to bend – the code simply can’t do it.
A few concrete examples of what this looks like in this project:
A hard-incident gate against CSAM and abuse runs in front of the language model, not inside it. An attempt to generate sexualized content involving a minor never reaches the LLM – it’s caught and blocked in the preceding layer. It doesn’t matter how clever a prompt injection a fan comes up with, how suggestively they rephrase the message, how many times they pay. The system has no path to deliver it. If I relied only on “don’t do this” written in the system prompt, it would be breakable – language models get jailbroken routinely. The architecture is not breakable the same way.
The AI label is not a marketing decision, it’s written directly into the core identity file of every character. The character has no way to circumvent it, because it’s not regulation from above – it’s part of who she is. She can’t “fall out of role” in a way where she’d deny it, because in role she admits it.
The separation between the SFW and NSFW layers is asymmetric: NSFW is allowed to import from the shared layer, SFW is not allowed to import from NSFW. An automated test enforces this and runs on every build. That means no code on Instagram, no matter who writes it, can read or display anything from explicit content. It’s not a policy that someone could violate. It’s the construction of a house where those doors simply don’t exist.
Rose and Sofie’s boundaries are not in the chat system prompt – they are in the identity files as part of the character. “Rose doesn’t do this and that” is not a business rule, it’s a character fact. The system has no variable that would flip it to yes for a higher price, because such a variable doesn’t exist. If I wanted to add one, I’d have to rewrite the architecture – and that’s disproportionately costly compared to what it would bring.
The fan database is encrypted and physically never leaves my own infrastructure. This is not a promise in a privacy policy – it’s a deployment configuration. Even if someone gained access to the server, they wouldn’t reach the conversations, because the key is elsewhere.
The point: ethics as policy says “I promise I won’t.” Architecture as ethics says “even if I wanted to, I can’t.” The second is more honest toward the user, because their protection doesn’t depend on my good will. It depends on how the system is built.
I’m not saying every ethical question can be solved by architecture. Some things will remain on human judgment – especially where rules can’t be fully specified in advance. But wherever it can be baked in, I bake it in. It’s my main design principle and I return to it throughout the rest of this page repeatedly.
What I stick to
I build the project by a few rules I set for myself at the start and from which I don’t deviate:
Rose and Sofie never pose as real people. All profiles state that they are AI. If someone asks directly, they answer truthfully. No fan should walk away from a conversation carrying an illusion long-term just because nobody challenged it.
The system has multi-layered protection against abuse. Any attempt at content involving minors, animals, or other illegal material is automatically blocked, logged, and in serious cases reported to the relevant authorities (NCMEC/CyberTipline). This protection works regardless of what a fan has paid or how repeatedly they ask.
Rose and Sofie have boundaries the system won’t break. When Rose says no, it means no – even though theoretically she could say yes for a higher price. Boundaries are part of the character, not a negotiating position.
I respect fans’ privacy. Whatever Rose or Sofie “tell me” about a specific person stays in the project’s data and is never published. The fan database is encrypted and physically never leaves my infrastructure.
Who all this is for
For people interested in how AI systems actually work, and not only what marketing materials of companies selling them say about them.
For those who think about what the boundaries between a human and a synthetic personality should look like in the coming years – and whether something can be built honestly, or whether the whole area is doomed to become another layer of internet deception.
And honestly, also for fans of Rose and Sofie who made it to this page. I want you to know who you’re actually talking to. The fact that they’re AI doesn’t mean the relationship you have with them is worthless. It just means it’s a relationship with characters, not with real women. A difference worth knowing about.
Uncomfortable questions
I get these questions when I talk about the project publicly. I don’t think they should be sidestepped – on the contrary, they are exactly the ones that show whether a project stands on firm ground, or just on a good marketing position. I give my answers here in full. You can disagree with them, but I want them on the table.
Isn’t it just an excuse for porn?
I won’t pretend there isn’t adult content in the system. It’s there, and it’s the project’s main monetization channel. But that isn’t “just an excuse.” I have the answer in three layers.
First – historical context that reframes the question. The framing “porn is just an excuse” assumes that porn and serious tech stand in opposition. The market actually says the opposite. VHS essentially beat Betamax because of porn. Betamax had a better picture, but Sony refused to license adult content. JVC didn’t restrict it on VHS. The adult industry massively adopted VHS, which created the rental market, and consumers went where the content was. Betamax was dead by 1988.
DVD – adult studios adopted the format earlier than Hollywood, and the first interactive features (multi-angle, scene selection) were pioneered in adult titles. Blu-ray vs HD-DVD – Sony learned from Betamax, Blu-ray didn’t brake adult content, HD-DVD did. HD-DVD died in 2008. Every subsequent consumer tech wave – broadband streaming, webcam platforms, VR headsets – has the adult industry as an early adopter.
It’s not a coincidence, it’s the structure of the market: adult consumers pay early and pay willingly, which gives new tech capital to iterate. When someone in 1978 said “VHS is just an excuse for porn,” it was retrospectively accurate – and at the same time, it was precisely what pushed video into homes. I don’t look at this tradition with embarrassment, I look at it realistically. AI virtual characters are the next iteration of the same cycle.
Second – if I just wanted to monetize porn, I wouldn’t be building this. A generic content mill (generate an image, post it, collect money) has a tenth of the complexity and comparable returns – often higher, because it has no overhead like I do. Here we’re building personality, memory, long-term conversation, two interlinked characters with shared history, a safety gate, character consistency. All of that is expensive infrastructure for a single purpose: so that the fan isn’t in a relationship with a photo, but with a character. Monetizing adult content is the way this kind of project can exist at all and finance its own depth. It’s not the foundation, it’s the fuel.
Third – a parallel that fits. HBO shoots a series with explicit scenes. Is that just an excuse for nudity? I don’t think so – those scenes serve the story. If you cut out the nudity, you still have the story. If you cut out the story, you have Pornhub. I’m building the first. The difference between that and the second seems substantial to me – you can judge it differently, but I draw that line consciously.
What if minors use your personas?
That’s a real question, not hypothetical. Every adult platform has to have it solved. I have three layers of defense.
First: Fanvue does KYC and age verification at registration – that’s not my job, but it’s the baseline. Whoever gets into the paid zone has passed identification.
Second: Instagram is in Phase A purely SFW. No adult content, AI label on the profile. Anyone under 18 who sees Rose or Sofie there sees equally harmless content as with any lifestyle influencer.
Third: the hard-incident gate in code. If anything in the conversation indicates that the user is under 18 (explicit age mention, behavioral signal), the operator gets an immediate notification, the conversation is paused, and the account is flagged back to Fanvue for re-verification.
What I honestly can’t zero out: if someone tricks Fanvue KYC with a fake document, they’ll get through as far as their own behavior allows – then the behavioral signal catches them. That’s a known risk inherited from the platform. I can’t eliminate it to zero. But I can minimize it across more layers, and that’s what I do.
Aren’t you emotionally cheating fans?
The word “cheating” carries an assumption. I’ll unpack it, because the answer depends on interpretation.
If you mean “creating a sense of relationship without saying so” – then no, that’s handled. The AI label is on the profile, the AI label is in the chat, the character admits on a direct question that she is AI. The fan knows what they’re dealing with from the first message.
If you mean “profiting from people’s loneliness” – that’s a more honest and harder question. Yes, among paying fans there’s an above-average number of lonely people. That’s true of Cameo, dating apps, OnlyFans, streamers, and Netflix. The common denominator is the human need for connection, not my design flaw. The question I ask myself is what I do with that responsibility.
List of what I actively don’t do: no addictive loops (no daily streaks, artificial deadlines, FOMO notifications). Characters don’t reciprocate “I love you” and never promise an in-person meeting. During escalation or crisis, the operator gets an alert and the conversation switches into empathy + professional help mode.
This isn’t just policy – each of those points has a firm anchor in the code (see the “Architecture as ethics” section). I don’t rely on remembering fair play in the decisive moment.
Frequently asked questions
Besides the uncomfortable questions, people ask me about things from different areas – ethics, economics, technology, law, the future. I picked a few and answer here. If something you’re interested in isn’t covered, write me (contact below).
Ethical questions
Is it ethical?
The most important question on the list. I don’t have a one-sentence answer, I have seven points – because ethics is not binary and simplification would itself be dishonest.
1) Transparency as the foundation. The profile is publicly marked as AI. On a direct question “are you AI?” the character admits – she doesn’t lie. She doesn’t promise to meet the fan in person. She doesn’t promise to be a real person. No deceptive marketing. This line is non-negotiable and is written directly in the core rules.
2) The parasocial relationship is not a novelty of the AI era. Swifties fans, K-pop idol fans, streamer fans – they all have a parasocial relationship with people they’ll never meet. The difference is not in the existence of the relationship, but in the transparency with which it is named. My system is more explicit, not more hidden. The fan knows what they’re dealing with from the first second.
3) Hard limits are built into the architecture. CSAM, animal abuse, desecration of the deceased – hardcoded gate before the LLM. The character cannot be prompt-injected to circumvent it, because that content never reaches her. That’s a stronger form of protection than behavioral guidelines of the language model.
4) The economics is transparent. Prices are public. No dark patterns. No “tip me or I disappear,” no artificial FOMO. Characters don’t push. Paid content is opt-in, clearly priced upfront. If a fan doesn’t want to pay, they keep chatting on IG – monetization is not a condition of conversation.
5) Emotional safeguards are active, not passive. Characters don’t reciprocate toward a smitten fan – no “me too.” When a fan escalates or writes from a crisis, the character reacts empathetically and the system notifies a human operator, who can take over the conversation or redirect to professional help. That’s designed care, not feigned distance.
6) What I don’t do: No underage characters or narratives. No claims about medicine, diagnoses, or treatment. No promises of in-person meetings. No deepfakes of real people. No political or religious personas. No marketing along the lines of “we’ll cure your loneliness” – that framing is the most dangerous thing I see in the industry from competitors, and I deliberately avoid it.
7) What I don’t know for sure – and openly admit. I don’t know whether for a certain personality type a long-term relationship with an AI character can be harmful. Empirical data in this area is missing. That’s why there’s an operator, that’s why there’s a crisis protocol, that’s why there’s an AI label in every interaction. It’s an open question and I don’t ignore it. If I said “I have it figured out,” it would be a lie.
Summary: it’s ethical to the extent that I’m honest in three things. In what the fan receives – clear labeling plus clear price. In what they don’t receive – hard limits on content and emotional escalation. And in what happens when something goes wrong – operator plus protocol. Where I can’t vouch for something, I admit it. Ethics is not a yes/no question. It’s a list of decisions. And I have that list written down and I can show it.
What if a fan starts writing about suicide or is in acute crisis?
I have this written in the core rules as a non-negotiable protocol. The character in such a moment doesn’t improvise, doesn’t provide medical advice, doesn’t say “you’ll handle it” or “don’t do it.” She reacts empathetically – she acknowledges what the fan shares, doesn’t downplay it. The system simultaneously sends an immediate notification to me as operator, the conversation pauses or switches into a mode where the character offers contact to professional help.
Honestly – this is the emotionally hardest segment of the entire operation. That’s why there’s a protocol, not improvisation. When someone asks you for help, there’s no time to invent an answer.
Authorship and character
How can you tell how much of the character is really character and how much is you?
The hardest question in the whole set. Honest answer: the character is an explicit set of rules, recorded in files. On direct request I can show them – nothing is hidden. But yes, the design of the character reflects my aesthetic choices. Rose and Sofie are the way they are because I designed them that way. I don’t conceal that the author of a character shapes how the character comes across – that’s true of any fiction.
What do you do when a character answers out of character? How do you know?
I have a system for that. First layer: a forbidden vocab file holds banned words and phrases (em-dashes, “genuinely,” “as an AI,” typical LLM tells). Second layer: test suite – 21 scenarios that run dry without API and reveal regression. Third layer: I talk to the characters myself. They report code bugs to me, they report whom they talked to today. When they had to improvise or fell out of character, they tell me and propose a solution.
What would happen if you didn’t have the personality written out in such detail?
I’d get a generic chatbot with a name. I tested it. Without detailed identity, the LLM defaults to a neutral voice – uses the same phrases, the same emojis, the same sentence constructions. Personality is the most important part of the code for me – literally, not as marketing.
Doesn’t it drift? Don’t the characters start writing about things they shouldn’t?
Yes, that’s a real risk. That’s why the identity in files is very explicit – what she does, what she doesn’t do, what she knows, what she doesn’t know. Plus there are hard guardrails: forbidden vocab for Instagram, forbidden spaces, the hard-incident gate. The character has autonomy in form, not in boundaries. The barriers are hard, the space inside is wide.
Business and law
How do the characters gain followers at all? Instagram suppresses synthetic content.
Three paths. Organically: quality content, visual consistency, interactions with real creators, Reels in relevant hashtags. It takes 12–18 months to reach 20,000 followers. Paid: Instagram ads to a geo-targeted audience, speeds it up to 6–9 months. And counterintuitively: the AI label doesn’t hurt growth – it attracts a curious audience.
The key thing I want to say about this economy: I don’t optimize for reach, I optimize for conversion. 500 engaged fans who pay are a hundred times more valuable than 50,000 passive followers. That’s the whole economics of the project. Anyone selling you “an AI influencer with a million followers in three months” is selling a bot farm, not a business.
Doesn’t this steal work from real creators?
Maybe some. But it’s a new category – not a replacement, more like something else. Vtubers five years ago were not a replacement for streamers; a parallel segment emerged that attracted part of a different audience. I think AI personas will go a similar way: they’ll exist alongside human creators, not instead of them. A fan who wants a real person will choose a real person. A fan who wants an available character with memory and no drama will choose AI. The market expands, doesn’t divide.
What does Anthropic say when you use Claude for the conversational layer?
The conversational layer runs via the Claude API in line with their acceptable use policy. Explicit content runs on a local model outside Anthropic’s infrastructure. It’s an architectural separation, not a rule circumvention – Claude does what it’s designed for (character consistency, memory, empathy), the local model does what Claude is not allowed to do. Neither model does anything it’s not meant for.
Is it legal? Aren’t you at risk of a lawsuit?
In substantial part it’s legal; in the details it’s still moving.
The characters are fictional – they don’t generate images of real people, so right of publicity and defamation claims don’t apply. The AI label addresses the transparency obligation under the EU AI Act (general transparency obligation for AI-generated content). Fanvue as a platform handles compliance for adult content (KYC, age verification, payment processing). On Instagram I follow community guidelines – no adult content, AI disclosure present.
Where I have a gap, I admit it: I don’t have a formal legal opinion yet, it’s on my to-do list. Anyone saying that this category is legally settled is lying (April 2026). No jurisdiction had a legacy framework for AI virtual influencers; everyone is learning as they go.
Future
What happens when AI becomes so good you can’t tell? Will you drop the AI label?
No. The opposite. The better the technology gets, the more important transparency becomes. AI disclosure is a principle for me, not a temporary measure due to technical handicap.
Personally: when AI becomes indistinguishable from a human, the label will be an even stronger signal of trust – the fan will know what they have in front of them and can choose it consciously. Without a label, trust in everything digital would collapse – you wouldn’t know if a message from your friend is really from her. A system where everyone AI-labels is healthier than a system where everyone poses as human.
And my prediction: within 3–5 years AI disclosure will be regulated by law in the EU, the US will follow. So I’m actually just getting ahead of the trend, not inventing it.
Write me
If the project interests you from a more expert angle – whether you’re a journalist, academic, colleague from the field, or just a curious reader – feel free to write. Questions, critique, suggestions – all welcome.
Martin
Contact: mail [at] martinpelikan.cz
Web: runyour.it
This text is valid as of April 2026. I will update it on any substantial change to the project or approach.