At Dexterity, we recently announced a strategic manufacturing partnership with Sanmina to scale Mech — our industrial superhumanoid. While the press release shares the “what,” I want to use this space to share the “why,” the “how,” and most importantly — the “what nearly broke us getting here.”
My first week at Dexterity, I visited one of our logistics customers in California. Coming from the semiconductor and solar world, I was stunned. I had spent years in Class 1 clean rooms — pristine environments where even a particle could ruin a wafer. Now, I was walking through a warehouse: dusty floors, forklifts whirring, cardboard everywhere, constant motion. And yet, there were our AI-powered robots — operating right in the middle of it all.
Two thoughts hit me immediately:
- I need to listen before I lead. This world is different. I couldn’t just copy and paste what I knew from the semiconductors and solar industries.
- The fundamentals still apply. Systems thinking, quality, manufacturability, reliability — all the lessons from semiconductors and solar still held. I just had to translate them into a new context.
Over the years, I’ve developed high-precision systems across complex industries. At Applied Materials, I ran engineering, applications development, and field operations. I led global teams, built teams in India, owned P&L for a product line, drove business transformation, shaped corporate strategy, and served as chief of staff. In startups, I’ve built supply chains from scratch, led new product introductions, and scaled manufacturing.
So, when we started developing Mech — our AI-powered, two-armed super humanoid designed to lift 60-pound packages and operate with human-like dexterity — I didn’t just see the need to prove the concept with a compelling demo. That part matters, of course. But I saw a deeper, more familiar challenge:
How do you take something that works once — and make it work every time, everywhere, at scale?
The Iceberg Is Real
When people see Mech lift 60-pound packages and gracefully navigate a warehouse like it was born there, it feels like magic. The motions are fluid. The interface is intuitive. It picks, places, adapts. It moves like it knows what it’s doing. And that’s the moment everyone sees — the working demo. That moment is the tip of the iceberg. But if you’ve ever built a physical product — one that blends AI, robotics, perception, real-time control, logistics-grade reliability, and safety-critical constraints — you know the truth: the demo is the easy part. The real test? Making it work every day. On every shift. With tired operators. On uneven floors. After a forklift bumps it. In heat, in dust. With zero room for failure. To build a Mech that actually works in the field — not once, but thousands of times — we had to engineer what lives below the waterline. And that’s where we’ve spent the majority of our energy. Let me paint a clearer picture. Under the surface of that magical demo, here’s what you’ll find:
Each of these layers carries cost, complexity, and risk. But ignoring them is not an option — because in the real world, these are the things that make or break your product. When the CEO of a major logistics company on a visit to Dexterity said “Mech works,” we smiled — because we know what it took to make that true.
Why This Matters: A Personal Lesson
Years ago, I worked at a solar startup trying to bring concentrated photovoltaic systems to market. We had a compelling concept. The early designs looked solid. We pushed hard — ramped up manufacturing, shipped units, deployed in the field. Everything was lined up. And then it crashed. Hard. Why? We hadn’t done the real work. We didn’t fully characterize the sub-systems. We didn’t model how the system would behave across extreme environmental conditions. We didn’t bring up manufacturing in a structured, process-controlled way.
What we missed in upfront engineering, we paid for — in time, money, and credibility. It took us three times longer and twice the budget to fix it. And we never fully recovered. That experience taught me three truths that I carry with me to this day:
- In hardware, you either pay up front or you pay a lot more later.
- Software development ends when it works. Hardware development begins when it works.
- Great hardware takes grit, scale, and discipline — far beyond the first demo.
At Dexterity, we chose to do things differently. We didn’t just commit to building a robot. We committed to building a product — one that could survive being shipped on a truck, scaled across warehouses, and serviced in the field by someone who wasn’t there for the prototype. That means sweating the details, every step of the way.
Startups Love the Surface. Physical AI Requires Depth.
In most software startups, you can ship fast, iterate, patch, and keep moving. Ship a demo, show traction, hit early metrics — and fix what’s broken later. That mindset doesn’t work in Physical AI. You can’t ship a half-built robot and expect to patch it in the field. I’ve seen it happen: teams with brilliant demos that completely collapse under the pressure of deployment.
Why? Because the invisible stuff — the “under-the-iceberg” details — will break you:
- Cable routing that adds two extra hours to every service call.
- A spec’d cable that fails after 500 bends because no one modeled the bend radius.
- A part that quietly goes end-of-life after your first 50 units.
- A pilot build that masks critical timing bugs between firmware and perception.
- A gripper mount with a torque spec that’s “good enough” — until it isn’t.
Any one of these issues can stall your deployment. Stack a few together, and you’re fighting fires instead of scaling. So we made a commitment: Design with the finish line in mind. Don’t build for the pilot. Build for the hundredth install. And the thousandth support call. That’s how you get from prototype to product — not with just vision, but with discipline.
What Makes Physical AI Different
In the world of robotics, the term “Physical AI” is still new — but it’s not hype. It’s a fundamental shift in how we build intelligent systems to solve real-world problems. At Dexterity, we’re not just developing software that thinks. We’re building systems that think and move — safely, reliably, and at scale. That sounds simple. It’s not.
In my previous roles at Applied Materials and in solar startups, I learned to think in systems. You couldn’t optimize a process without accounting for hardware limitations. You couldn’t scale a product without understanding manufacturing constraints. And you couldn’t survive without balancing cost, quality, time, and reliability. The same applies here. Maybe even more so. Because in Physical AI, we deal with atoms, not just bits. Every improvement in the system has a cascading cost:
- Change a motor driver? That might require requalifying the entire drive module.
- Move a sensor? That shifts the perception pipeline, the calibration model, and the control loop.
- Push a firmware update? That bug could brick a $200 actuator before the watchdog even notices.
Unlike pure software, you can’t just roll back an update. Unlike traditional robotics, intelligence isn’t static — it’s learning, adapting, retrying, and improving. But intelligence without reliability is just chaos. That’s the core challenge of Physical AI:
You have to build systems that are both:
- Smart and rugged,
- Flexible and dependable,
- Cutting-edge, but field-ready.
And that’s what makes this work hard — and important.
Dexterity’s Systems Strategy: Why the Product Is the Moat
We often get asked: what’s your moat? It’s not just the AI models. It’s not just the robot arm. It’s not just the user interface. It’s the system.
Our strategy borrows from companies like Apple and Android. Not in how their products look — but in how they think:
- A tight, integrated product that evolves rapidly without breaking compatibility.
- A platform that gets more valuable as more customers use it.
- And applications — built by us or partners — that solve real business problems.
At Dexterity, this strategy comes to life in three core layers:
• Product. Our robots are architected for rapid innovation, using common, commodity components where possible. We focus our own energy on areas of strategic differentiation — our gripper system, our force-torque sensing, our onboard server. The parts of the system that make Mech superhuman.
• Platform. We’ve built an AI + data platform that learns over time. The more Mechs we deploy, the smarter and more efficient they become — not just individually, but collectively. It’s a network effect for physical systems.
• Applications. This is where everything comes together: workflow-driven business solutions that solve specific pain points — like palletization, truck loading, or truck unloading — built with the right balance of software, robotics, and human-in-the-loop control.
This system-level thinking is what allows us to outmaneuver siloed competitors. Some companies build good software. Others build strong robots. We build both — and more importantly, we build them to work together. That’s our moat:
- Not a single feature, but the full stack.
- From the sensor to the cloud. From the pick to the pallet.
What Comes Next: From Strategy to Execution
The systems approach — integrating AI, robotics, perception, and software into a tightly orchestrated product — is what gives Mech its edge. But building a smart system is only the beginning. The real challenge lies in translating customer needs into engineering requirements that guide every design choice. In the next post, I’ll go deeper into the engineering mindset behind Mech: how we convert real-world constraints into architecture, how we balance tradeoffs, and how we build with discipline from day one. Because the system only works if the engineering behind it is sound.
