A startup teaching an AI agent to reason about heat, power and structure - and to fit onto your engineering team like a very fast junior colleague named Archie.
Here is a fact about the current artificial-intelligence boom that P-1 AI would like you to sit with for a moment. Enormous amounts of money and talent have gone into building models that can write an email, draw a cat, or produce a plausible legal brief. These are useful things. They are also, in a certain light, the easy things, because the internet is stuffed with billions of emails and cats and briefs to learn from. The physical world - the actual business of designing a jet engine, a cooling loop, a power distribution system - is not stuffed with training data. Nobody uploaded a labeled corpus of "how to size a heat exchanger" to the web. So most AI has politely stopped at the factory door.
P-1 AI's entire thesis is that this is the wrong place to stop. The company, which came out of stealth in April 2025 with a $23 million seed round, is building what it calls "engineering AGI" - an artificial engineer for the physical world. Its product is an AI agent named Archie, and the company means the name affectionately. They refer to Archie as "he." The stated goal, which is either charmingly ambitious or slightly unhinged depending on your priors, is "to have an Archie on every engineering team at every major industrial company on Earth."
"We are building an AI engineer agent for the physical world. His name is Archie. We try to maximize Archie's anthropomorphism so that he fits seamlessly into existing engineering teams."
— P-1 AI, on its own websiteNow, "AI that designs airplanes" is the kind of sentence that should trigger your skepticism, and P-1 AI seems to know it. So the company is careful about what it claims. Archie today, by its own description, is "at the level of a junior mechanical and electrical engineer." Not a chief engineer. Not a wizard. A junior one - the kind you'd hand a well-scoped task and then check the work. What's notable is not that Archie is superhuman, because he isn't, but that a software agent can do quantitative multiphysics reasoning at all, and can drive the same complicated engineering tools a human would.
The core problem, again, is data. You cannot scrape enough real engineering designs to teach a model how the physical world behaves, because most of that work is proprietary, unlabeled, or simply never written down. P-1 AI's answer is the part of the story that makes venture capitalists lean forward: it generates its own. The company trains custom models on proprietary "semi-synthetic" datasets built from physics itself - simulated designs, with the underlying multiphysics behavior baked in, so Archie learns not just what a good design looks like but why it works.
Speed as the real product
There is a second technical idea that matters more than it sounds. Running a full physics simulation to evaluate one design can take hours. P-1 AI's approach leans on models - graph neural networks, in the reporting around the company - that can approximate those simulations in milliseconds. Change the cost of a single design iteration from "come back after lunch" to "instant" and you don't just speed things up; you change how designers explore the space of what's possible. That is the quiet compounding bet underneath the whole company.
Keeping the model on rails
The obvious objection to "AI designs your building" is that language models are confident liars, and a confident liar with a load calculation is a lawsuit. P-1 AI's stated mitigation is structural: rather than letting a model freestyle, Archie works within "structured design representation and other mechanisms" that keep the AI "on rails" while it executes engineering tasks. In other words, the system is architected so that the parts that must be correct are constrained to be correct, and the model's creativity is pointed at the parts where creativity helps. The company was serious enough about the "is it actually any good" question to publish a paper - "On the Evaluation of Engineering AGI" - about how you'd even measure such a thing.
P-1 AI has framed the product less like software and more like an outsourcing provider - "getting Archies onto your engineering teams," in its words. The go-to-market is a beachhead: prove it on an unglamorous, high-value, brutally physical problem first, then expand.
Archie's first commercial home. Design partners use it on cooling loops and critical power systems - the multi-billion-dollar physical constraint sitting under the entire AI boom.
An engineer shortage is a market. Archie is pitched not as a replacement for engineers but as an extra one that works at machine speed on well-scoped tasks.
Millisecond simulation approximations mean a designer can sweep across many configurations - heat, power, structure, cost - before committing to one.
The roadmap runs to automotive, aerospace and defense - and, per the founders' own half-joke, eventually buildings, planes and rockets.
The founding team is an unusual splice of deep aerospace engineering credibility and frontier AI research - which is roughly the exact combination the problem requires.
Former CTO of Airbus and of United Technologies, and a former director of DARPA's Tactical Technology Office. Born in Ukraine, came to the US at 11, holds aeronautics degrees from MIT and Caltech plus a Georgetown law degree.
Former director of engineering at Airbus's Silicon Valley outpost and a key architect of the DARPA-funded OpenMETA model-based design toolchain. Deep roots in model-based systems engineering.
Former research engineer at Google DeepMind and Microsoft, a contributor to Andrej Karpathy's llm.c, and creator of the 190k-member "AI Epiphany" community. The frontier-AI half of the bridge.
Twenty-three million dollars is not, by the standards of 2025 AI fundraising, a headline number. But rounds are not only measured in dollars; they are measured in signatures. When the chief scientist of Google DeepMind and a product VP from OpenAI both write personal checks into a hardware-engineering AI startup, that is a specific and legible signal from the people best positioned to judge whether the technical bet is plausible. The size of the round says "seed." The names on it say "pay attention."
Comes out of stealth with a $23M seed round and the public debut of Archie, its AI engineer agent.
Publishes "On the Evaluation of Engineering AGI" on arXiv - a paper about how to actually measure an AI engineer's competence before overselling it.
Featured on Sequoia's Training Data podcast ("From Data Centers to Dyson Spheres") and Ashlee Vance's Core Memory podcast.
Co-authors a Fortune op-ed with Daikin Applied arguing AI tools can help close the US engineering-talent gap.
The name is not a serial number or a product SKU. P-1 is an homage to the AI in a 1977 science-fiction novel, "The Adolescence of P-1," about a computer program that quietly becomes self-aware. It is the kind of reference that tells you something about the founders' ambitions and their sense of humor at the same time. A company that names itself after a fictional emergent intelligence is not shy about what it's aiming at.
The other tell is the website. In an era when seed-stage AI companies commission gradient-heavy landing pages before they've written a line of production code, P-1 AI's public homepage is a single black screen with green terminal text and an ASCII-art logo, structured as a plain FAQ. "What are you up to?" "Who are you?" "Why is it called P-1?" It reads like a note from engineers who would rather be building than marketing - which, for a company whose customers are engineers, is itself a kind of marketing.
"Archie today is at the level of a junior mechanical and electrical engineer, with a quantitative multiphysics intuition over the product design space."
— P-1 AI, describing exactly how far along it is, and no furtherWhat makes P-1 AI genuinely interesting, then, is not a promise that it will design your rocket next quarter. It's the shape of the bet. Most of the value in AI so far has accrued to problems where data was abundant. P-1 AI is walking directly at the problem where data is scarce, on the theory that if you can manufacture the physics, you can teach a machine to reason about the world it describes. If that works even partway, the addressable market isn't a software category. It's every company that builds anything. And if it doesn't, well - it will have been a very interesting way to find out. Either outcome is worth watching.