The useful thing about an obsession is that it saves time on introductions. Ethan Kinnan's personal website gets to the point in nine words: “A human with AI can do anything.” Beneath it sits a live tally of his age, rendered to four decimal places, because apparently even a birthday should behave like a dashboard. The joke contains the operating system. Kinnan likes feedback, motion and questions that lead to something he can run.
That disposition has carried him from financial-data internships to Lindy, then into a two-person San Francisco company called Sherpa. The company studies how people move through a website, finds where they hesitate or leave, builds a new version and tests the change on live traffic after approval. A brochure, in this telling, is an unfinished experiment wearing nice typography.
Kinnan and his co-founder, Norbu Sonam, did not discover the problem on an alpine retreat with color-coded sticky notes. They had worked together at Lindy, the AI-agent company, and watched website experiments crawl through the familiar queue: idea, analysis, design, engineering, quality assurance, launch. Each participant was sensible. The system was slow. Their idea was to compress the handoffs into software.
The internship automation department
Kinnan studied finance at the University of Washington's Foster School of Business from 2020 to 2024, while adding applied mathematics and courses in probability, linear algebra and machine learning. He had already won Washington State DECA's financial-operations research event in high school. Finance supplied the material; automation became the recurring extracurricular.
At KSH Advisors, PitchBook Data and Standard Metrics, he gravitated toward repetitive processes and tried to make them disappear. When GPT-3.5 arrived, he built an AI parser for financial documents. He later attempted an AI financial planner with two close friends. On his website, he records the outcome without varnish: the project failed, and the trio intends to return. Failure receives a salute emoji and a place in the chronology. It does not receive a funeral.
The projects became stranger and more ambitious. After Claude Sonnet 3.5 arrived, Kinnan began building a new web application every day. Working as an AI Automation Engineer for Allie K. Miller, he built a data analyst that suggested changes from metrics and a self-researching system that evaluated more than 100 multi-agent frameworks daily. The young finance graduate was learning software by repeatedly asking software to help him make software. Recursion can be an excellent teacher when it also ships homework.
Move toward the useful people
Kinnan graduated without a job lined up or a perfect account of what came next. His later description of the period is refreshingly procedural: work somewhere he could use AI all day, learn quickly, move to San Francisco and surround himself with people closer to the future he wanted. He chose inputs over prophecy.
At Lindy, he became the company's first forward deployed engineer, a job built around closing the distance between a product and the people attempting to use it. He helped customers build agents, taught more than 10,000 people in live sessions and moved into ecosystem and partnerships work. In March 2025, he announced that he had moved from Seattle to San Francisco. He liked the density of the place. “You can't really escape it,” he said of the city's AI conversation, adding that his obsessive personality enjoyed being in the middle.
The move also brought him alongside Sonam, an engineer who had worked on Microsoft Copilot before building agent infrastructure and running growth experiments at Lindy. Their skills met at an unusually practical seam. Kinnan understood customers, education and distribution. Sonam understood the technical machinery and the experiment queue. Both had felt the same irritation.
The old employer becomes customer one
Sherpa began in 2026, with Lindy as its first customer. The arrangement had an elegant economy: the founders knew the problem, the buyer knew the founders, and the website supplied a busy stream of real behavior. In an early interview, Kinnan said Sherpa increased Lindy's total conversion by 20 percent in the first weeks. The later launch account reported 40 experiments in 30 days and a 30 percent increase in total site conversion. The figures describe different windows, but the direction gave the pair a reason to continue.
The product proposes more than fresh headlines. It can build layouts, animations, pricing treatments and landing pages matched to an ad. Sherpa's current site divides the work into three modes: improve existing pages through experiments, create new pages, and adapt what appears according to a visitor's context. The company's larger bet is that the page should be less like a monument and more like a competent shopkeeper, attentive to who entered and what confused them.
Human approval remains part of the mechanism. Sherpa says a proposed version is shown before publication and can be reverted. That detail matters. “Self-improving” makes a fine slogan; a checkout page inventing its own prices at midnight makes a fine crisis. The product's choreography gives the machine the tedious parts and leaves the consequential click with a person.
A shark with manners
A clever feedback loop does not exempt a founder from sales. Kinnan has been candid that Sherpa's first five customers were difficult. The team had begun too passively. His corrective was vivid: “You really just have to be a shark and talk to people and ask people for help.” They went through their contacts, sent roughly one to two hundred messages, and looked for introductions, potential customers and honest feedback.
The shark, notably, comes with manners. Kinnan pairs the request with a belief in exchange: another person's time is valuable, so offer something useful in return. Some people say no, others vanish, and neither response deserves an opera. Continue. His advice to prospective young founders is equally plain: use the tools intensely, take the leap and remember that the field of people who actually try is smaller than it appears.
That mixture of velocity and persistence shows up in his side projects. For a short film called Teddy, he tested Runway's Gen-3 Alpha model by sending a teddy bear through time. When the bear's face changed slightly between clips, Kinnan gave it a red bow tie to distract the eye and create the appearance of continuity. The detail is playful and revealing. A new tool rarely behaves perfectly. Put a bow tie on the seam, learn what the model can do, and make the next thing.
From conversion to self-improving software
Sherpa joined Y Combinator's Spring 2026 batch and describes itself with a crisp instruction: make websites improve themselves. Conversion is the first proving ground because the outcome is measurable. A visitor arrives, acts or leaves, and the experiment produces a number. The team's longer horizon is broader. Kinnan and Sonam imagine websites, applications and products continuously adjusting from real usage, whether the visitor is a person or an AI agent.
The agent part is an intriguing wrinkle. As software agents increasingly research products and make recommendations, a company's website has two audiences: people reading the pitch and machines parsing the evidence. Sherpa says it can test agent-facing versions to learn which information helps a company get recommended. The old art of presenting oneself has acquired a second critic, and this critic reads structured pages at machine speed.
Wins Washington State DECA's financial-operations research event and begins at the University of Washington.
Builds finance and AI projects, enters DubHacks Next and graduates from the Foster School of Business.
Moves to San Francisco, joins Lindy as its first forward deployed engineer and later leads ecosystem work.
Co-founds Sherpa with Norbu Sonam and completes Y Combinator's Spring batch.
For now, the company remains small enough for every result to be personal. A test wins, a customer replies, a message goes unanswered, a page changes. Kinnan's career has accumulated through the same rhythm. Automate one irritating task. Build one app. Teach one room. Move one city. Ask one more person. The method has no grandeur until you look backward.
Golf, his favorite sport, offers the tidier metaphor. You choose a line, take the shot and receive immediate evidence about your judgment. The scorecard neither flatters nor argues. Kinnan is building a company around a comparable idea: let behavior edit the plan. It is a patient philosophy delivered at startup speed, which may be the only pace he finds amusing.