Stella: MVP in 2 Days, $350K/Month in 2 Months — an AI App Built by a Solo Creator With 4 Million Followers

Dollar figures are converted at an approximate rate of $1 = ¥150.

The story of “I built an MVP in two days” isn’t especially rare anymore, post-generative-AI. What’s rare here is that 1.2 million and 2.8 million followers were already waiting on the other side of those two days.

Starter Story’s report on Stella, a smartphone AI manifestation/visualization app, puts monthly revenue at $300,000 (about ¥45M) two months after launch, and $350,000–360,000 (about ¥52.5M–54M) by the time of the interview. There are 12,000 paying customers and over 200,000 downloads.

Looked at through the numbers alone, this reads as “an AI app that hit.” But following founder Sarah Pearl’s own account (she goes by “Hot High Priestess” on social media) reveals a sequence that runs opposite to typical indie development. She didn’t build the product and then go looking for distribution. She finished building distribution, then slotted a product inside it.

The published numbers

What it sells

Users enter their desires or goals, and a video/audio piece is generated in which a “future self who has already achieved it” speaks to them, along with a daily affirmation. Meditation and visualization content already exists in abundance, but the problem Pearl identified was one specific gap: it wasn’t personalized for the individual user. Her view is that generative AI made it possible to solve exactly that one gap at nearly zero marginal cost.

Implementation runs on generative-AI-native development tools like Cloud Code, Google AI Studio, and Cursor. She isn’t an engineer herself. “The technical barrier to entry has disappeared,” she says flatly.

The decisive move happened “before building”

The most concrete part of Pearl’s account is her validation process. While still just conceiving the app, she posted short-form videos on the exact problem the app would solve, and read the comments. “You get comments like ‘I wish something like this existed’ or ‘this is what I struggle with.’ That’s the blueprint,” she says.

In other words, her demand validation wasn’t a survey or user interviews — it was the response from her existing distribution network to a video she’d already thrown out there. “Everyone builds front-to-back. They make a product, then go looking for distribution. I do it backward, I build the content and distribution first, then the product.”

The effect of this sequence shows up in the numbers. At the moment of launch, the audience to announce to already existed: 1.2 million on Instagram, 2.8 million combined on TikTok, over 1 billion cumulative views. It’s not disclosed how many of the 200,000 downloads came from existing followers, but the fact that 12,000 paying customers accumulated within two months with no mention of paid advertising shows that most of the customer acquisition cost was absorbed inside the existing distribution network.

Why this structure is fast

The obvious reason is that validation cost drops to near zero. A video, if it hits, becomes proof of demand. If it misses, all that’s lost is production time. Compared to the sequence of building first and then running A/B tests, the loss on a miss is an order of magnitude smaller.

The less obvious one: the language of demand is supplied by users themselves. What’s written in the comments becomes the ad copy, and the feature spec, directly. The step of guessing what to build disappears.

The deepest reason is that launch stops being a single one-time event. Pearl says: “We rebuild the formats that worked, over and over.” Repeating the pattern of a hit video means distribution functions as an ongoing acquisition engine.

She’s also specific about “what didn’t work.” “Even videos that mention Stella tend to convert worse if they don’t show the actual usage moment.” What works is showing the screen in use rather than name-dropping the product, an observation with real granularity. She treats critical comments the same way, as algorithmic fuel: “the most successful videos were the most controversial ones.”

What’s reproducible, and what isn’t

What isn’t reproducible is obvious. A distribution network of four million people isn’t a result of Stella. It’s a precondition for it. And that network sits on top of the money and track record of a TikTok Shop brand doing 7 figures a month, launched a year and a half earlier. An individual developer aiming for the same initial velocity is unlikely to land the same outcome. The timing, when AI personalization was still novel, also won’t return in the same form.

What is reproducible: the sequence itself, surface demand through content before building, holds structurally even with an audience of just a few thousand followers. The insight that videos showing actual usage convert better isn’t scale-dependent either. On development speed too, her flat statement that “speed is the number one ingredient for success in this industry” reflects a market structure where the window before you get copied is short.

Caveats when reading the numbers

What’s published is MRR (monthly recurring revenue), not cash received, not profit. App Store/Google Play fees (15–30%) and refunds haven’t been deducted. The $300,000 and $350,000–360,000 figures come from different statements within the same interview, and the article title uses the former.

More importantly, retention and churn are entirely undisclosed. Manifestation-style apps tend to run hot in month one, with retention past month three being hard to predict. There’s no guarantee the two-month MRR figure scales cleanly to an annual run rate around $4M.

And Pearl herself says: “this app category is full of copycats.” If generative AI lets you build an MVP in two days, that’s just as true for your competitors. This business’s defensive moat sits entirely on the distribution side, not the product side. Her repeated advice, “invent something that solves a problem that hasn’t been solved yet”, is likely the flip side of that same awareness.

Related reading

• PhotoAI — Pieter Levels’ solo-run operation — another case of an AI product run solo, with a contrasting approach to defensibility.
• PDF.ai — Damon Chen — for contrast, an indie build that started with no existing distribution network.