Aurorin CAD Is Quietly Ripping Out the 1980s Engine Inside Every CAD Tool You Know
A Y Combinator W26 startup rewrote the geometry kernel that SolidWorks and its peers still lean on. The payoff it's chasing: a part that takes 20 minutes to model, made in seconds - or just described out loud.
Ask a mechanical engineer about their CAD software and you will hear a particular kind of sigh. It is the sigh of someone who has waited for an assembly to open, watched a single edit cascade into a dozen broken features, and paid a few thousand dollars a year for the privilege. Aurorin CAD, a startup in Y Combinator's Winter 2026 batch, has built its entire pitch around that sigh - and around a theory about where it comes from.
The theory is uncomfortable for the incumbents. Aurorin argues that the reason professional CAD feels slow is not your laptop, your graphics card, or your patience. It is the geometry kernel - the mathematical core that every modeling tool uses to represent solid shapes. Most of those kernels were written in the 1980s. Aurorin wrote a new one.
The part that took twenty minutes
The company's headline claim is deliberately concrete: a part that takes an experienced SolidWorks user 20 minutes to model can be made in seconds in Aurorin. There are two ways to get there. You can use a fast, modern editing interface built for current CPUs and GPUs. Or you can describe the part in plain language to an integrated AI agent and watch it produce real, editable solid geometry - not a picture of a part, but the actual parametric object an engineer can keep working on.
Company figure. The gap Aurorin is selling is felt in the gut, not read off a spec sheet.
That phrase - "real, editable solid geometry" - is the whole game. Plenty of tools can generate a mesh or a render from a prompt. Very few can hand you a parametric feature tree you can then edit by hand. Aurorin's bet is that the difference comes down to how deep the AI reaches into the software.
The product itself is a single desktop application, not a suite. Inside it are the two workflows engineers actually toggle between: the direct-manipulation interface, where you sketch, extrude, fillet and constrain the way you always have, only faster; and the chat panel, where you offload the tedious parts - "add four M4 mounting holes on this face, 8mm from each edge" - and keep the result as a feature you can revise. Aurorin's stated design goal is that neither mode feels like a demo. The interface is meant to stay responsive on complex parts and large assemblies, which is precisely where legacy tools tend to bog down.
Why the kernel is the story
Here is the part most people skip. Underneath SolidWorks, Siemens NX, PTC Creo, CATIA, Autodesk Fusion 360, and Onshape sit a small number of geometry kernels - most famously Parasolid, which dates to 1988, and ACIS from the same era. These kernels handle the hard, invisible math: boundary representation, numerical precision, topological correctness, performance at scale. They are genuinely impressive pieces of engineering. They were also written for hardware that looks nothing like a 2026 workstation.
Aurorin's founder, Michael Baron, decided not to license one of those kernels and build a friendlier interface on top - the path almost every CAD startup takes. Instead he rebuilt the kernel itself. It is the hardest possible starting point, and it is also the only one that makes the AI story work.
"The agent is working with the same geometric primitives the kernel exposes natively, at the same level of abstraction an experienced engineer would use manually."On why AI has to reach the kernel
Read that again with an engineer's skepticism. When an AI bolts onto legacy CAD through a plugin, it is talking to the software from the outside, guessing at operations the kernel was never designed to expose cleanly. When the kernel is yours, the AI can speak its native language. That is what Aurorin means by "AI-native," and it is why the company shorthands itself as "Claude Code for mechanical engineers" - a coding assistant, but for solid geometry.
Who is actually waiting for this
The clearest customer is the mechanical engineer at a hardware startup. They are the ones most likely to feel the pain sharply: they move fast, they iterate constantly, and they resent paying roughly $4,000 a year per seat for software that stalls when an assembly gets big. Aurorin's public asks are pointed - hardware companies that want faster prototyping, and design engineers willing to send feedback. It is a narrow, thin market on purpose. Thin markets are where incumbents pay the least attention.
Aurorin's own framing of the status quo is blunt: "It's common for existing software to take 4 hours to open CAD assemblies." Whether or not that is your worst case, the direction is the point. The people who lose hours to loading bars are exactly the people who will try something new the moment it is free.
There is a second, quieter pain Aurorin names: the fragility of parametric models. Edit one early feature and downstream features can break in a cascade, forcing an engineer to hunt through a broken feature tree. Anyone who has rebuilt a model after changing a single dimension knows the feeling. Aurorin lists that cascade explicitly among the problems it is trying to design out - alongside slow performance and interfaces it describes as inherited rather than chosen.
The business, the model, the moat
Aurorin is free to download for Mac and Windows today, which is the whole go-to-market in one move. In a market defended by procurement contracts and IT approval cycles, the fastest route past the gatekeeper is the engineer who already likes your tool. Get the individual hooked, then sell teams and enterprises later. It is freemium, but aimed at a place where freemium is unusual.
This is the part that separates Aurorin from a wrapper. Anyone can ship a chat box and a freemium plan in a year. A production-grade B-Rep kernel is different work - the kind that takes years and a rare flavor of computational-geometry talent. That difficulty is the barrier. It is also the reason CAD has been called a graveyard for startups: the technical bar is brutal, and the incumbents have decades of procurement relationships. Onshape, one of the more ambitious recent attempts, was acquired by PTC for a reported $470 million - a genuine outcome, but also a reminder of how the story usually ends.
The founder in the engine room
Michael Baron's background reads like a deliberate warm-up for this problem. He was a three-time SpaceX intern who worked on Raptor engine combustion simulation, Dragon guidance, navigation and control, and Starshield flight software. He also worked on GPU driver performance at Apple. Rocket combustion teaches you numerical precision under pressure; GPU drivers teach you to squeeze modern hardware. Both are exactly the muscles a from-scratch CAD kernel demands.
"A part that takes an experienced SolidWorks user 20 minutes to make only takes seconds in Aurorin."The company's core claim
It would be easy to over-read the pedigree. What matters is that Baron chose the unglamorous layer. He did not start with the 3D viewport or the marketing site. He started with the geometric core that the entire industry treats as solved and untouchable. Whether that bet pays off is unknown - the company is small, early, and walking into a market that has swallowed better-funded attempts. But the shape of the wager is unusually clear for a seed-stage company.
Where it fits
Aurorin sits at the intersection of three things happening at once: hardware startups multiplying, AI agents getting good enough to do structured technical work, and a generation of core engineering software running on foundations older than the engineers using it. Most "AI for X" companies paint over the third problem. Aurorin's answer is to rebuild the foundation so the AI has something real to stand on. If it works, the interesting outcome is not a slightly better SolidWorks. It is a design tool where the distance between describing a part and holding an editable model of it collapses to almost nothing.
For now, the honest summary is smaller and truer: a two-person team in San Francisco, $130,000 from Y Combinator, a new kernel, and a free download. The claims are big. The company is not - yet. What it has is a clear theory of why CAD feels the way it does, and the nerve to fix the layer everyone else leaves alone.