Why Invisible Technologies Exists: The Stealth Startup Probl

There's a company that claims it has trained more than eighty percent of the world's leading AI models. Cohere. Microsoft. AWS. And in twenty twenty-two, OpenAI hired it to help fix ChatGPT's hallucinations. Most people have never heard its name.

Okay, that's a weird sentence. Eighty percent? And no name recognition?

The name is Invisible Technologies. Which... is either the best or worst branding decision in Silicon Valley, depending on the day.

They committed to the bit.

Founded October twenty fifteen by a guy named Francis Pedraza. Cornell grad. And his first company was a total failure — an app called Everest, a personal goal-tracking thing, backed by Peter Thiel. It shut down in twenty fourteen because users just... stopped coming back.

What's his takeaway from that?

He told Inc. magazine, joking: "Have a business model."

Fair.

So Invisible starts as a ten-thousand-dollar-a-month virtual chief of staff service for CEOs. He wrote later that they thought nobody would sign up. They got five clients and a waitlist, raised five hundred thousand dollars... and the model collapsed within six months because they couldn't actually do the work efficiently.

So it died.

No — it mutated. From April twenty sixteen to August twenty seventeen they worked with exactly ONE client. One. And in that time they built over fifty internal processes and something they called the Digital Assembly Line.

Which is what, in English?

Break a company's workflow into tiny steps. Pedraza calls them Legos. Automate every Lego you can. And where software chokes — the messy edge case, the judgment call — you drop a human in. A global network of subject-matter experts.

Hang on. That's... a staffing agency with software on top. Why is that a two-billion-dollar company?

That is the exact objection he got from investors. And here's the turn: when he pitched, VCs told him "services" was a dirty word. They wanted recurring revenue, high margins, infinitely scalable from day one. He didn't get the money.

And he... ignored them?

He decided the pushback WAS the opportunity. If every investor flinches at the same word, that's a blind spot, not a flaw. So he kept building the unfashionable thing. Quietly. Profitably.

Profitably? That's not a sentence you hear in AI.

Twenty twenty-one, profitable. And then two moments made them. March twenty twenty: DoorDash calls. Pandemic's starting, restaurants are flooding onto delivery, and somebody has to type in every menu and every price. Invisible's own case study says they sped up onboarding by two hundred thirty-three percent.

Humans typing in menus. The glamorous AI frontier.

And then twenty twenty-two. OpenAI hires them for reinforcement learning from human feedback on what becomes ChatGPT — specifically to cut down factual errors. Hundreds of contractors, titled "advanced AI data trainers."

What did the job actually look like?

You read conversations between the model and users. You flag anything incorrect, illegal or offensive. You score responses one through seven — factual accuracy, grammar, harassment. And then you write what you think the ideal answer would have been. That goes to OpenAI.

So the thing that feels like magic when you type into a chatbox is partly... a stranger grading essays on a scale of one to seven.

That's the whole thesis of the company, said out loud. Pedraza did something else unusual too — he borrowed against the business to buy out the early venture investors. As of February twenty twenty-five, about three hundred current and former employees held fifty-five percent of it.

The employees own the majority. That's genuinely rare.

Then January twenty twenty-five, new CEO: Matthew Fitzpatrick, who ran QuantumBlack Labs at McKinsey, overseeing a thousand engineers. Pedraza stays chairman. And on September sixteenth, twenty twenty-five, they announce one hundred million dollars in growth funding, led by Vanara Capital — a brand-new firm spun out of TPG. It was Vanara's first public deal.

What's the valuation?

Bloomberg reported more than two billion, citing people familiar. The company itself didn't put that number in its press release — it said total capital raised, one hundred forty-four million, and twenty twenty-four revenue of one hundred thirty-four million, more than double the year before.

So a ten-year-old company only raised a hundred and forty-four million total. That's... restrained.

That's the point. And Fitzpatrick's pitch is blunt: enterprises are spending vast sums on AI and few are making it work. Invisible's own site claims as many as ninety-five percent of AI projects fail — because companies drop models into messy systems without doing the human work first.

So why stay stealth for a DECADE if the story's this good?

Because the story only became good recently. For most of that decade, "we put humans inside the automation" sounded like an admission of failure. Now it's the product. In March twenty twenty-six they bought WeCP, a technical-assessment platform, to verify that their experts are actually experts. In January twenty twenty-six they joined the World Economic Forum. Fitzpatrick's line, January fifth: humans will stay involved in data creation for decades.

So the invisibility wasn't modesty. It was waiting for the market to catch up.

The company's own about page says it best, almost smugly: the companies that win won't be the ones that use AI. They'll be the ones that stopped needing to talk about it.

Which is a strange thing to put on a website.

Every time a model feels effortless to you, somebody was paid to make it look that way. The whole industry's built on a layer nobody photographs. Invisible just named itself after it.