Hestus has raised a $7.4 million seed round to build what it calls mechanical intelligence: AI that understands geometry and the manufacturing constraints behind it. The round was led by Insight Partners, with Chalfen Ventures, Rock Yard Ventures and Garage Capital participating. Together with an earlier $1.5 million pre-seed, the San Mateo company has now raised $8.9 million.
Its first product is an autocomplete for CAD software. As an engineer sketches in CAD, Hestus proposes the next move (a constraint, a dimension, a geometry) and the engineer accepts it with a single keystroke. We spoke with co-founder and CEO Sohrab Haghighat about why he believes the real bottleneck of the physical AI boom sits on the engineer's screen, why he is not building on a large language model, and how 5,000 engineers adopted the product through purely organic growth.
Hestus at a glance
- Company: Hestus, Inc.
- Headquarters: San Mateo, California, USA
- Founders: Sohrab Haghighat (CEO) and Kevin Chu (CTO)
- Accelerator: Y Combinator, Summer 2024 batch
- What it does: AI copilot inside CAD software that reads the engineer's design intent and proposes the next steps
- Product: Sketch Helper, available for Autodesk Fusion on Mac and Windows; SolidWorks integration in private beta
- Users: More than 5,000 engineers
- Latest round: $7.4M seed, announced October 2, 2026
- Lead investor: Insight Partners
- Other investors: Chalfen Ventures, Rock Yard Ventures, Garage Capital
- Pre-seed: $1.5M led by Liquid2 Ventures and Rock Yard Ventures, with Y Combinator, Kyle Vogt and Eric Migicovsky
- Total funding: $8.9M
The founders: hardware engineers who did the clicking themselves
Haghighat and Kevin Chu were among the first ten employees at Cruise, the self-driving car company. Haghighat went on to run SpaceRyde, a space launch startup, while Chu led engineering at the medical device company Mendaera. Both spent their careers building physical products, and both know where the hours go.
"Kevin and I, we come from the hardware world. We are hardware engineers," Haghighat says. "We have suffered from the pain of manual clicks, and all the details that you need to juggle in your head. Our ultimate goal is to make engineers' lives easier."
That background is also his filter for deciding what to build. "I always ask myself: back when I was doing my studies, back when I had my first job, or back when I was running SpaceRyde, would a tool like this have been helpful for me? If the answer is yes, we have to build it. If the answer is no, we should not do it."
The bottleneck is on the engineer's screen
Haghighat frames the problem from two directions. The first is physical AI. If robotics keeps advancing at its current pace, he expects an explosion in the number of robots the world will need to manufacture. Most of the attention, and most of the money, goes to the AI that lets those robots understand the world. "But before we can actually build a physical product, you need to design that physical product," he says. Over the next five years, demand for physical goods will rise without any matching increase in the capacity to design them.
The second direction is the tool itself. Every physical object, down to the water bottle on your desk, started as a digital design in CAD software. "The CAD that we have been using hasn't really changed in forty years. That's not necessarily a bad thing," he says. "The problem with CAD is that it doesn't understand the one thing that it needs to understand. That is manufacturing."
So every design goes through rounds of iteration before it can be built. "There is a huge gap between a fancy 3D model and a 3D model that is ready to get manufactured," he explains. And that gap, in his view, is what the reshoring debate keeps missing. "Everyone thinks the real bottleneck is manufacturing capacity. They're trying to build more factories, bring manufacturing back to the US, back to Europe or Canada. But the real bottleneck is going from the concept in the head of the designer to a finished, manufacturable design that we can pass to the factory. And that requires the engineers to juggle lots of factors in their heads and manually implement them in the CAD."
The company puts a number on it: of the total work on a design, only 20 to 30% is the creative work of conceptual design. The rest is documenting that design in CAD, which is the part Hestus wants to automate.
Autocomplete today, mechanical intelligence tomorrow
Sketch Helper reads the sketch as the engineer draws it. Draw a rectangle and it offers both centerlines; place a circle in one quadrant and it can propose the mirrored set with the symmetry constraints already applied. Suggestions appear as a red overlay, the engineer cycles through them, and one key accepts. The company says users work 2.5 times faster with four times fewer clicks.
When I compared this to what happened in software, where AI tools gradually learned the constraints of real production code, Haghighat agreed on the trajectory and used it to describe his roadmap. "Coding agents started from very, very baby steps: single line completion, variable name suggestions. Then it got better, it could write a for loop. Then a function. Then it could write software, but it missed the security checks and made stupid mistakes. And now it is getting better," he says. "We want to go through the same steps of progression."
Today, Hestus predicts one step ahead. Next, it should propose ten steps ahead. After that, it should handle a full part or a collection of parts, and then bring in manufacturing cost optimization: "How can we make this, still as good and functional, but cheaper, so we can mass manufacture it?"
Each step requires a model that understands the physical world. "What does it mean for two gears to properly engage? For a tolerance to be loose, or requiring a press fit? That is ultimately what we want to build, what we call mechanical intelligence."
Why Hestus is not building on an LLM
Here Haghighat parts ways with the software analogy. "What is powering the revolution in software development is LLMs. However, we strongly believe large language models are not necessarily the best architecture to solve the physical design problem."
Instead of fine-tuning a language model to reason about geometry, Hestus is designing its own architecture from scratch. He points to precedents. "No one uses an LLM to generate pixels for an image or a movie. Image and video generation are primarily powered by diffusion-based models. It's a different architecture." A model that is purpose-built for spatial-reasoning, he argues, can be far more compact and more capable at reasoning about physical and manufacturing constraints. "We are not trying to solve the AGI problem. We are only trying to solve the mechanical intelligence problem."
A smaller, specialized model also has a commercial argument. "The use of AI is skyrocketing. Now everyone is thinking: do I have the budget for this? How many tokens can I afford? If you come up with a model that is more efficient in its token consumption, of course it is going to be well received by its users."
Text for specifications, the mouse for shapes
Hestus' website makes fun of the long text prompt an engineer would need to describe a simple sketch to a chatbot. Haghighat is careful to say the company is not against text. Text has a role: carrying specifications. A business team comes back from customers with requirements for the next product. Regulators such as the FAA, FDA or NHTSA add their own standards too. "The best way of communicating that to an AI is via text, via uploading the document."
Shapes are another matter. "If I ask a mechanical engineer to describe a shape with words, now that's backwards. 'The circle that is to the right and on top of that line, not the three millimeter one but the four millimeter one, move it ten millimeters.' No one does it." Often the next part simply has to fit the geometry that is already on screen. "You just load the CAD, and an AI that is trained properly can see the pattern. I don't need the engineer to explain it to me. As soon as they start working, I probably know their next ten steps. Let me go and propose it to them."
In his view, many AI CAD startups exist because founders saw an opening after coding agents or legal AI took off, and asked which frontier came next. "I claim the majority of people running those companies have never spent a single day designing a physical product, or set foot in a machine shop." Hestus' approach is to learn from the progression of coding agents and rebuild each step for the way mechanical engineers work. "Asking a mechanical engineer to work like a software engineer, or the other way around, I believe it's the wrong approach."
Go-to-market: let the engineer be the champion
Hestus reached 5,000 engineers without targeting companies. "This was all a bottom-up approach, going through our network, getting it into the hands of each individual engineer and letting them be the champion of the product. It was purely organic growth."
The model, again, is how coding tools grew. "It wasn't a big company adopting it. It was each individual engineer adopting it, using it for personal projects, introducing it to their colleagues, and over time making their way into smaller companies first, then medium-sized, then bigger organizations. We believe the same path is what we need to follow."
Haghighat learned this lesson from the buyer's side, while running SpaceRyde. Another founder pitched him a tool for engineering workflows, he liked it, and he asked his team to try it. A few weeks later, the most outspoken engineer came to his desk: why was he forcing them to use it? The team had already built those pieces internally six months earlier. "That was when I learned that the demand for an engineering tool should come from the engineering team. It should not come from upper management."
From then on, the pattern reversed. Engineers came to him asking for software, always with the same argument: remember that milestone where we lost so much time? This tool would solve the same problem five times faster. They would agree on a trial and a success criterion, and if it worked, buy a short-term subscription. "That's how it should be. The engineer should be the champion for the tool. The manager should not push it top down."
When I suggested this also builds the eventual enterprise sale, with hundreds of a company's engineers already using the product by the time anyone talks to the economic buyer, he agreed. But for now, the focus stays on users. "If I want the engineer to be the champion of my tool, I need to get on the same level and ask them: what's not working for you? What are the top three, top five challenges stopping you from doing the work you went to school for and love doing?" Hestus does not always solve the problem, he admits. "100% not. But when they see that we moved the needle, and that we cared enough to listen and develop something for them, you build a relationship with your user. That feedback, once you work on it, is the most important ingredient for success."
In practice, the company wants to be known by every single mechanical engineer: by mechanical engineering students, by engineers at startups, and by engineers at large enterprises. The next distribution step is SolidWorks itself: the integration is in private beta, available on request, and a public release roughly doubles the number of engineers Hestus can reach. More CAD platforms will follow.
Still, Haghighat sees the product as the main growth channel. "Every new user has certain requirements, certain use cases. We learn from those and build more features, and hopefully those features are valuable to other people, which attracts other users. Every day we ask ourselves: what feature is going to be most appreciated by our engineers?"
What the $7.4M is for
The largest share of the round goes to hiring, on two fronts: senior and staff-level software engineers, and machine learning researchers. "We have a big vision and a big dream that we want to fulfill, but we cannot do it all by ourselves," Haghighat says. Readers with that profile can find open roles on the Hestus careers page.
His long-term ambition goes beyond saving clicks. "Everyone's body is different. As we get older, we might need prosthetics or assistive devices, and something that fits my body may not fit someone else's. What if new devices could be custom designed for each person?" The same logic applies to an aircraft designed in half the time and still more fuel-efficient, a new car, or medical robots that take decades to design and build. "What if we can shrink the time for these products to come to market?"
GTM lessons for deep tech founders
- Name the bottleneck others overlook. Hestus positions itself against the common belief that factories are the constraint, and places the problem between the concept and the manufacturable design.
- For engineering tools, demand comes from the engineer. A tool pushed by management gets rejected; a tool an engineer asks for gets a trial, a success criterion and a budget.
- Follow the adoption path of developer tools. Individual users first, then colleagues, then small companies, then larger organizations, with enterprise deals arriving once usage is already there.
- Fit the interface to the work. Text for specifications and requirements, direct manipulation for shapes. A chat box is not the default answer for every profession.
- Live inside the tools users already have. A direct integration with Fusion and SolidWorks means nothing new to learn and no files to move, and each new integration opens a new user base.
- Listen even when you can't fix everything. Asking users for their top problems, and coming back with progress, builds the relationship that keeps product-led growth going.
