He builds flight simulators. Not for pilots. For the AI agents learning to do the jobs inside your company's software.
Lupo, photographed in the frame most founders live in now: a square avatar, a camera-aware half-smile, the whole company staked on the sentence he's about to say next.
Somewhere on a server rented by Deeptune, an AI agent is opening a Salesforce record it will never see, inside a company that does not exist, to complete a task no human assigned. It gets the task slightly wrong. It gets a small negative reward. It tries again. This is the product. Tim Lupo, the 20-something co-founder and CEO of Deeptune, sells the simulated Tuesday - a high-fidelity replica of an accountant's or a support rep's or a DevOps engineer's workday, run thousands of times so that a model can practice being useful.
Lupo describes it more plainly than the category deserves. "We essentially build simulations of digital work that look like the workspace of an accountant or a lawyer or a software engineer," he told Fortune when Deeptune announced a $43 million Series A in March 2026, led by Andreessen Horowitz. The industry has a jargon word for these things - reinforcement learning environments. Lupo prefers the gym metaphor, or better, the cockpit one.
"You wouldn't have a pilot who has only ever read books or watched tutorials fly a plane," he says. "You would put them in a flight simulator. What we build are essentially the flight simulators for AI doing work across the economy." It is a good line because it contains the whole thesis. Today's models can pass exams and recite procedure. Ask one to actually run a multi-step process across real software and it stumbles, because it has read about the work without ever doing it. Deeptune sells the doing.
The bet has a specific shape. Static training data - the scraped internet - is running thin, an industry problem sometimes called data exhaustion. Deeptune's answer is to convert the collection of experience from a labor problem into an engineering-and-compute problem: build the environment once, then let models generate their own training signal by trying tasks millions of times and getting rewarded when they succeed. The company says it has built hundreds of these gyms, and that they've already contributed to the recent jump in AI "computer use" - the shift from answering questions to operating software.
What makes Lupo worth watching is not that he arrived at this idea. It's that he got there second, from an entirely different business, and turned the whole company around to chase it. Deeptune is on its second life. The name is the only thing that survived.
Here is a detail that would be easy to bury and shouldn't be. In November 2023, Tim Lupo and his co-founder Lukas Schmit landed on the Forbes 30 Under 30 list - in the Social Media category. They were 24 and 25. The company was called Deeptune, and it did something quite specific and quite different from what it does now: it used AI to dub creators' videos into multiple languages while preserving their original voices. A tool for reach. A creator-economy play.
That Deeptune is gone, and the current one kept the letterhead. The reinvention runs deep enough that the a16z announcement and the company's own blog describe a reinforcement-learning infrastructure firm with no trace of dubbing. The through-line is Lupo's comfort with the machinery of AI itself - the same instinct that made voice cloning interesting made simulated experience interesting, one cycle later, when the field's bottleneck moved from content to capability.
Founders talk about pivots as if they were graceful. Most are not. A pivot means telling investors, employees, and yourself that the thing you were recognized for is not the thing you should be doing. Lupo did it with the confidence that now colors everything he says. "We were the first company to build an environment a bit over a year ago, and no one really knew if it was going to work," he told Fortune. "We now know that they work insanely well."
"Insanely well" is a founder's phrase, and normally you'd discount it. The benchmark numbers make it harder to. On OSWorld, a test of AI computer-use, frontier models have now edged past the human baseline - the kind of result the environment-builders quietly feed.
OSWorld measures whether an AI can actually operate a computer to finish real tasks. The line to beat is a human. Recent frontier models cleared it - the kind of gain that better training environments help produce.
"We essentially build simulations of digital work that look like the workspace of an accountant or a lawyer or a software engineer."
"We were the first company to build an environment a bit over a year ago, and no one really knew if it was going to work. We now know that they work insanely well."
"If you want to be in New York and work on frontier AI or AGI, Deeptune is one of only a couple places you could join - and probably the only early-stage place you could join."
"What we build are essentially the flight simulators for AI doing work across the economy."
The default map of frontier AI has one city on it, and it isn't New York. Lupo built his company there anyway, in person, and turned the contrarianism into a recruiting pitch. "If you want to be in New York and you want to work on frontier AI or AGI," he says, "Deeptune is one of only a couple places you could join, and probably the only early-stage place you could join."
The roster suggests people are buying it. The roughly 20-person team is drawn from Anthropic, Scale AI, Palantir, Hebbia, Glean, and Retool - the kind of pedigree that usually decamps to the Bay. Lupo's own network is part of the draw. a16z, in its announcement, credited him with "a rare combination of technical depth and product intuition" and close relationships with top AI researchers, which is investor-speak for: he knows the people whose models he is trying to improve, and he ships what they actually need.
That intimacy shows up in the cap table too. The angel list on the Series A reads like a who's-who of the exact researchers and operators who would know whether RL environments matter: OpenAI's Noam Brown, Mercor CEO Brendan Foody, Applied Compute CEO Yash Patil. When the people building the models write personal checks into the company that trains them, the signal is hard to misread.
Lupo frames the stakes in the largest available terms. He calls the work among the most critical in the pursuit of AGI - simulating whole sectors of the economy so that model capability turns into real, measurable performance on the job. Grand, and possibly true, and either way the sentence he'll keep saying until the next raise.
Deeptune reconstructs a real software workflow - the tools, the steps, the mess - so an accountant's or support rep's day exists as a sandbox.
Models attempt assigned tasks inside the simulation over and over. Lupo calls each attempt a "rollout."
Correct behavior earns a virtual reward; wrong turns don't. The signal is high-quality and generated at scale.
Reinforcement learning turns thousands of tries into competence the model couldn't get from static text.
Instead of scraping more internet, models manufacture their own experience - a compute problem, not a labeling one.
The environments go to frontier AI labs, where they've fed measurable gains in computer-use capability.
His pinned GitHub project is a JupyterLab extension for reviewing pull requests - and his profile wears an "Arctic Code Vault Contributor" badge.
Deeptune's environments helped push AI models past the human baseline on the OSWorld computer-use benchmark.
The company kept its name through a total reinvention - from AI voice dubbing to reinforcement-learning infrastructure.
Before founding Deeptune, Lupo was a founding engineer at the AI search company Hebbia.
The Series A closed on the same New York in-person team Lupo insists is a feature, not a compromise.
USC full merit scholarship - a distinction the school reserves for a fraction of a percent of applicants.
Sources: Fortune, Andreessen Horowitz, Deeptune, SiliconANGLE, Forbes, LinkedIn, Crunchbase, GitHub. Figures and quotes drawn from public reporting on Deeptune's 2026 Series A.
Tim Lupo is the co-founder and CEO of Deeptune, a New York startup that builds high-fidelity reinforcement learning environments - which he calls 'training gyms' or 'flight simulators' - where AI agents learn to do real digital work across enterprise software like Salesforce and Slack. In early 2026 the company raised a $43M Series A led by Andreessen Horowitz. Deeptune started life as an AI video-dubbing tool that landed Lupo and co-founder Lukas Schmit on the Forbes 30 Under 30 list, then pivoted to become RL infrastructure for frontier AI labs. A former founding engineer at Hebbia and a full-merit USC scholar, Lupo pitches the roughly 20-person, in-person team as one of the only early-stage places in New York to work on AGI.
Last updated: