Ballet barre to code review. The founder teaching AI agents how to fail safely before they go live.
Here is the setup: one founder, trained for fourteen years in pre-professional ballet at Boston Ballet, San Francisco Ballet, and Ballet West, who simultaneously won three YoungArts awards for photography, researched self-assembling robots at Columbia, published simulation work in IEEE out of Stanford's Robotics lab, became the first hire at a generative AI startup, watched it get acquired by Salesforce, and then decided to build the infrastructure that every AI lab on earth quietly needs but nobody had built yet. That is Nicolai "Nic" Ouporov, Co-Founder and CEO of Fleet AI.
Fleet builds reinforcement learning training environments - "RL gyms" in the company's own terminology - where AI agents can practice operating real-world software applications under realistic conditions before anyone turns them loose on actual work. Think of it as a flight simulator, except the pilot is an AI agent and the cockpit is a replica of Salesforce or Excel. Fleet creates these simulation environments so that labs can train models on real, consequential tool-use without the consequences. The company went from roughly $1M in annualized revenue at the end of 2025 to over $60M in early 2026 - a 60x increase in a matter of months.
Before Fleet, Nic was building in the same space from a different angle. At Stanford's Robotics and Embodied AI Lab, he studied how simulation environments could help robots learn to work alongside humans - the same fundamental question, pointed at physical machines rather than software agents. At Columbia's Creative Machines Lab, it was physics simulations and self-replicating robots. The thread running through all of it: how do you build a world inside a computer that teaches an intelligent system how to behave in the real one?
Respell, where Nic was founding engineer, gave him his first direct look at the commercial side. The startup built no-code generative AI automation tools and was acquired by Salesforce in January 2024. That proximity to how enterprises actually use AI - and where agents break down - fed directly into Fleet's thesis. The problem was not that AI agents lacked capability. The problem was that they had no safe place to develop judgment.
Nic co-founded Fleet with Fred Havemeyer in 2024. The company's total funding has reached $45M+, with a Series A led by Craft Ventures and participation from Sequoia Capital, Menlo Ventures, and SV Angel. As of April 2026, Bain Capital Ventures is leading negotiations on a new round targeting at least $50M at a post-money valuation of approximately $750M. Fleet employs roughly 40 people and is headquartered in New York, with Nic based in San Francisco.
The name Fleet has a double meaning that Nic has cited explicitly: a team of agents operating collaboratively toward a shared goal, and the ability to move quickly. Both describe what the company is building and how it operates.
"I try to lead my artistic endeavors as a sort of scientific inquiry."- Nicolai Ouporov, Ratrock Magazine
The idea behind Fleet is straightforward and slightly obvious in retrospect: AI agents need to practice before they go to work. The analogy Nic has used is medical residency - doctors spend years in supervised clinical settings before operating unsupervised on patients. Fleet gives AI agents the equivalent: high-fidelity simulation environments where they can operate replicas of production software, encounter edge cases, fail without consequence, and learn.
Fleet's core product is RL gyms - reinforcement learning training environments that replicate popular enterprise applications like Salesforce and Excel at sufficient fidelity for agents to actually train on them. AI labs use these environments to develop models that can operate complex software toolchains. The simulations are designed to expose failure modes that only emerge through repeated, varied use, not through static evaluation benchmarks.
The market context matters here: as agentic AI systems move from demos to production, the question of "does this agent know how to behave?" becomes critical. Evaluation benchmarks answer that question after the fact. Fleet's infrastructure answers it before - through practice. Nic's background in robotics simulation gave him an intuition that the same physics of teaching machines through experience could be ported from physical robots to software agents.
The revenue velocity - $1M to $60M ARR in under a year - suggests the problem Fleet is solving was already urgent when they arrived. A negotiating round at $750M valuation, led by Bain Capital Ventures, with Sequoia and Menlo as follow-on investors, confirms the market agrees.
* Bain Capital leading new round negotiations as of April 2026
"Artists are, in many ways, entrepreneurs."- Nicolai Ouporov, Columbia University
Nic Ouporov discusses Fleet AI's mission and infrastructure for training AI agents