He built a personal AI therapist that nobody wanted. The code underneath it became the infrastructure for one billion phone calls.
In the summer of 2023, Jordan Dearsley walked the streets of San Francisco talking to a voice AI he had built for himself. He called it Harmon. It was an AI therapist — not a company pitch, not a product roadmap. Something he actually needed.
To find out if anyone else might find it useful, he did what resourceful founders do when they have no distribution: he walked onto buses and airdropped digital contact cards to strangers, asking them to call his AI. Most didn't. The ones who did weren't looking for therapy. They were developers who wanted to know how he had built something that actually sounded like a real conversation.
That gap between what Dearsley made and what people wanted turned out to be worth more than the original idea. Within two years, those early voice conversations became Vapi — the infrastructure platform now handling one to five million calls daily, trusted by Amazon Ring, Intuit, New York Life, and a million-plus developers around the world.
"When we first built Vapi, I had to airdrop contact cards to random people on the bus to get them to call it. Last week, the Vapi team stood in front of 2,000 builders and watched someone lose their mind calling 1-844-HEY-VAPI."— Jordan Dearsley, on X (@jordan_dearsley)
The leap from bus rides to sold-out developer conferences did not happen through product-led growth alone. It happened because Dearsley and co-founder Nikhil Gupta — both University of Waterloo software engineering dropouts who met as roommates — understood something most voice AI builders got wrong: the models were never the bottleneck. The infrastructure around them was.
Dearsley has been candid about what the early days of dropping out actually looked like. He and Gupta relocated to San Francisco after being accepted to Y Combinator on what felt like a whim — they had applied with a school project idea and genuinely didn't expect to get in. Finding their footing in an expensive city while burning through runway meant going to soup kitchens at times. The sacrifice, Dearsley has said, pushed them to think bigger. Small bets didn't feel worth the cost of making them.
Before Vapi, before Superpowered, before Y Combinator, Jordan Dearsley ran the second-largest fidget spinner store on Etsy. It is a detail that sounds like a joke until you consider what it reveals: the instinct to spot a market, move fast, and build a channel. That same instinct later showed up in how he found Vapi's first users — not through press, but by physically putting the product in front of people and watching what happened.
Superpowered, the company Dearsley and Gupta built through Y Combinator's W21 cohort, was not one product. It was a dozen bets in sequence.
A lecture platform. A single-button meeting joiner that reached ten thousand weekly active users. An AI-powered note-taker that rode the remote-work surge and grew to roughly half a million dollars in revenue using Deepgram's transcription technology. Each iteration taught them something; none of them felt like the right thing to be building for the next decade.
Then, in mid-2023, came the Zoom call that changed everything. It started at 11 PM with Michael Seibel, one of Y Combinator's most respected voices. Dearsley has described it publicly as a conversation that was honest to the point of uncomfortable — the kind of feedback that resets your entire frame. After three years of productivity tools and calendar software, the question was no longer "how do we grow this?" but "is this the problem we actually want to solve?"
The answer was no. Dearsley burned out on productivity apps. He wanted something technically harder. He moved back to the drawing board — which, in his case, meant walking through San Francisco and talking to an AI he built for himself.
"The most significant thing was making a big sacrifice just pushed us to think larger — we didn't want to work on anything small anymore."— Jordan Dearsley
Most voice AI startups compete on which LLM they use. Dearsley decided that was the wrong question. The models, he argued, were already good enough — and getting better without his help.
What nobody had solved was the layer in between: the real-time audio routing, the sub-500ms latency requirements, the orchestration of transcription and speech synthesis and language models firing in sequence, the telephony integration, the guardrails that keep an enterprise call from going off-script, the observability tools that let a VP of Customer Experience know exactly what their AI agent said and why.
Dearsley has framed Vapi's entire mission around this insight: AI models are capable but unpredictable. Enterprises — the ones routing 100% of their inbound call volume through your platform — require predictability above almost everything else. Vapi's moat is not a proprietary model. It's the operational layer that makes any model trustworthy enough to put in front of a customer.
When Dearsley says "we're everything in between the models and actually talking to customers on the phone," he means it literally. Vapi handles the plumbing that nobody wants to build but everyone needs: real-time audio transmission, model chaining, conversation flow logic, tool calling APIs, and the after-call analytics that help enterprises iterate.
His estimate: only about 10% of companies genuinely benefit from building a custom voice stack. For the other 90%, the time investment is impossible to justify. And for those companies — Amazon Ring, Intuit, New York Life — Vapi became the answer after a long search.
Amazon Ring's evaluation alone tells the story. Ring considered more than 40 AI voice vendors before choosing Vapi. Once selected, Ring moved from zero to full production deployment in two weeks. One hundred percent of Ring's inbound calls now route through Vapi. Customer satisfaction scores improved.
"The golden problem is taking this indeterminate beast that is a model and taming it. If you can do that, then you can provide value to the world."
"We're good at infrastructure, building good developer experiences and good product experiences. So we just decided to double down on that."
"The real unlock is building agents for your customers that feel human."
"I don't think anyone that's not running real-time audio systems should be running real-time audio systems."
"We're everything in between the models, regardless of the state of them, and actually talking to customers on the phone — and everything that comes after that."
"That latency piece was really our differentiator at the time — the reason people would use us rather than roll it themselves."
He is not comfortable with small. The soup kitchens, the dozen pivots, the decision to walk away from half a million dollars in ARR — all of it points to someone for whom the cost of the wrong problem is higher than the cost of starting over.
Dearsley has described the hardship of the early Superpowered years not as a badge of suffering but as a forcing function. When you've dropped out of school, relocated across a continent, and are eating at a soup kitchen in San Francisco, you don't have the luxury of being precious about your product. You ship, you talk to users, you pivot. The fidget spinner store on Etsy years earlier was the same instinct — find the market, move fast, don't overthink the category.
The detail that captures his approach best might be the walking. When he built Harmon, the AI therapist, he walked two hours a day testing it. That's not a launch strategy. That's a founder who genuinely uses his own product because he needs it — and whose clarity about what doesn't work comes from lived experience rather than a user research report.
His honesty about that period — and his willingness to talk about it publicly — is one reason the developer community around Vapi is as engaged as it is. People build with founders they trust. Trust comes from the ones who don't pretend the messy parts didn't happen.
Hear directly from the founder on voice AI, developer infrastructure, and building through uncertainty.
"AGI is here. Why am I still on hold?"
This is Vapi's central question — and the headline on Dearsley's Series B blog post. The gap between what AI can do and what enterprise customer service actually feels like is a business opportunity worth billions. Vapi is closing that gap one call at a time.
Dearsley's stated target market is the $400 billion global call center industry. His frame for it is not "replace call center agents." It's more specific, and more ambitious: make every customer-facing voice interaction self-improving.
Today's enterprise contact center is largely stuck in the IVR era — press 1 for billing, press 2 for support — despite the fact that the consumer experience of voice AI (Siri, Gemini Live, ChatGPT Voice) has moved far beyond it. That gap, Dearsley argues, is not a technology problem. It's a deployment infrastructure problem. Enterprises haven't been missing better models. They've been missing a reliable, observable, controllable layer to put those models into production at scale.
Vapi's roadmap runs toward enterprise reliability: tighter guardrails, deeper observability, predictable behavior in high-stakes workflows, and multi-agent orchestration that lets complex customer journeys span multiple AI handoffs without losing context. The word Dearsley uses most consistently is not "smart" — it's "reliable."
"The real unlock is building agents for your customers that feel human."— Jordan Dearsley, Vapi Series B announcement
When Kavak — one of Latin America's largest used car platforms — deployed Vapi agents across sales, operations, and support, revenue in the Mexico market doubled. When Instawork shifted candidate interviews to Vapi agents, available worker supply doubled. These are not marginal productivity gains. They are structural changes to how a business operates.
That's what Dearsley means when he says he wants to turn contact centers from cost centers into revenue centers. And it's why, after evaluating 40+ vendors, Amazon Ring called Vapi.