He tweeted a screenshot, put up a single HTML page with a Stripe payment link, and fulfilled the first 100-200 orders by hand overnight at $30 each — roughly $400-500K in the first month or two before any of it was automated
$30 per order · 100-200 orders fulfilled manually on night one (~$3,000 for one night's work) · fine-tuning cost $3 per training at launch, raised to $20 by the supplier mid-flight (founder-reported)
Fewer bars = easier, cheaper, or faster for an AI-assisted solo builder. Editorial judgments based on the case details.
The chain of accidents matters, because it is the actual method. Pieter Levels started playing with image models like everybody else in late 2022, typing prompts. He noticed the models produced beautiful architectural houses, so he shipped a toy site — thishousedoesnotexist.org — that generated random house designs. Working on that, he noticed the models also produced beautiful interiors, so he built InteriorAI, where you generate interiors and then upload a photo of your own room and have image-to-image modify it. That worked. Then he experimented with fine-tuning — taking the base model and training it toward a narrower goal. He trained on interior photos and the results improved. So he tried training on photos of himself, "just to see what happens," and got pictures of himself in every conceivable style. His reaction was simply "wow, this is very cool," so he tweeted it. The tweet went viral. His words for what happened next: "the next day, I was like, I need to really quickly make a startup for this. So I made avatar AI." He believes it was the first big AI avatar startup. What he built that day was not software. It was a single HTML page showing example avatars and the input photos that produced them, plus a Stripe payment link. Orders arrived immediately — around 100 of them — and he fulfilled them by hand: read the Stripe notification emails, collect the customer's photos from a Typeform, upload them to the fine-tuning service, wait, download the results, send them back. He did 100 to 200 orders this way himself, spending an entire night on it. At $30 an order that is roughly $3,000 for one night of manual labour. He started automating on day two and had it fully automatic after about a week. The reported total: somewhere around $400,000 to $500,000 in the first month or two.
6 more sections — the full story, the playbook, the risks, and the numbers. Free account unlocks everything.
Create free accountEmail code only. No password.The quick-reference card — acquisition channels, replication playbook, and risk map — unlocks with a free account.
Create free accountEmail code only. No password.Paste into Claude Code / Codex and get a working version of this product end-to-end.
Two production-grade prompts per idea: a full build spec and an SEO growth plan. Copy, paste, ship.
Go Pro — $10/moCancel anytime.Every claim traces back to a source video.
Source video links unlock with a free account. Full transcripts are Pro.
Create free accountEmail code only. No password.Data credibility: Founder interview on The Bootstrapped Founder podcast, first person, no dashboards or screenshots available in the transcript. The revenue figure is deliberately loose in the founder's own telling: "I think I made like, maybe half a million dollars or $400,000, lots within a month or so — I'm getting two months." We report the range and the ambiguity rather than pick a number. The $30 price point is confirmed in a direct exchange with the host, who then computes ~$3,000 for the first night's 100 orders. The manual fulfilment process (index.html page, Stripe payment link, Stripe checkout redirecting to a Typeform for photo uploads, working the queue from Stripe notification emails, 100-200 orders done by hand, automation started day two and completed in about a week) is described in detail and is the most reliable part of the account. The supplier price increase from $3 to $20 per fine-tune, the subsequent move to Replicate, the ~$40M reportedly made by VC-funded clones, and the claim that this was the first big AI avatar startup are all founder-stated and unverified — the $40M competitor figure in particular is hearsay about a third party. Twitter follower count is given as both 300,000 and 340,000 at different points in the conversation. Region is Netherlands on the basis of the founder saying outright "I'm Dutch"; note that he is a long-term digital nomad who describes splitting the year between southern Europe and Southeast Asia, so the business has no fixed operating country. The product-origin chain (thishousedoesnotexist.org → InteriorAI → fine-tuning experiments → AvatarAI) is as he describes it.