Subscribe
Share
Share
Embed
Deep conversations with the founders, investors, and operators building real-world AI - robotics, automation, industrial systems & AI infrastructure. Past the headlines, into how these technologies are really built, deployed, and scaled. Hosted by Bogdan Cristei, venture partner and former systems engineer.
The Blindness of Modern Manufacturing | Jared O'Leary, SirenOpt
The OPTIM Update
---
Bogdan Cristei (00:00)
This is the Optim Update. I'm Bogdan Cristei and today's guest is Jared O'Leary, CEO and co-founder of SirenOpt. SirenOpt is a deep tech company building what they call manufacturing intelligence — the ability to see what's actually happening to a material while it's being made, in real time, on the production line. If you've ever wondered why even the most advanced factories in the world still can't fully control their processes, this conversation is going to make that make sense. We talk about the physics, the commercial reality, and what it's actually going to take to build an autonomous factory. And without further ado, here is my conversation with Jared O'Leary.
Bogdan Cristei (00:41)
Mr. Jared, welcome. Excited to have you on. How are you feeling? Are you ready to talk about manufacturing intelligence and autonomous factories?
Jared O'Leary (00:49)
Yeah, absolutely.
Bogdan Cristei (00:50)
So let's start. You said something to me that I keep coming back to. You said that manufacturers today are essentially more or less blind. What did you mean by that?
Jared O'Leary (01:01)
When I say that, I'm really talking about companies making very advanced products and advanced materials — things like batteries, semiconductors, and so on. What I mean is that manufacturers have an extreme amount of control over what their machines do, and a lot of insight into the environments those machines operate in. They also have a lot of insight into the performance of the end product. But what they really miss — where they have very limited insight — is the quality of the product as it's actually being made. They miss the moments where material properties shift, defects form, and yield is lost. They can see clearly when they lose yield or when a product doesn't meet spec, but they miss the actual moments where that happens.
Bogdan Cristei (01:51)
Interesting. Got it. So before we even get into what you're building, help me understand the landscape a little bit. When people say advanced manufacturing, what are we actually talking about? Because I think most people picture something like a car assembly line.
Jared O'Leary (02:28)
Yeah, I think people even imagine something like a Model T — a bunch of people with screwdrivers and hammers assembling parts. But when I say advanced manufacturing, I'm talking about semiconductors, batteries, jet engines, even modern car engines. The general theme across almost all of what I consider advanced manufacturing is that there are very thin and very sensitive layers of different types of materials. A material could be a metal, a coating, a composite of different tiny bits in different form factors. If you want one overall theme — what separates manufacturing a hammer from manufacturing a battery — it's this: a hammer is what you see. There's a head and a handle. But if you look inside a battery pack, there are literally thousands upon thousands of unique material layers, each with slightly different chemical compositions, different thicknesses, different densities, different porosities.
Bogdan Cristei (04:11)
Got it. These days we're seeing a lot of money flowing into reshoring and domestic manufacturing. What's actually driving that, and why does it matter for the kind of manufacturing you're focused on?
Jared O'Leary (04:24)
I think we should rigorously define reshoring and what that money is actually going into. A lot of companies right now — big automotive companies — are raising money and claiming they're reshoring manufacturing and putting plants in America, but in reality, Chinese companies are putting their employees, with Chinese equipment, in Chinese plants, and just sticking an American company name out front. People hear the term reshoring and think America is just going to magically start manufacturing all these things, but these ecosystems are far more global and interconnected than that. The general goal is really just getting manufacturing to occur on American soil. Why does that matter? One, there are geopolitical constraints — we saw recently how actions in the Strait of Hormuz affected oil prices, but also caused the price of helium gas to skyrocket. Helium is an important industrial gas used in ways most people never think about. So there's an incentive to have stability against unpredictable geopolitical events. And second, manufacturing creates high-paying, sophisticated, stable jobs — not Model T assembly line jobs.
Bogdan Cristei (06:19)
Back to the blindness analogy. I want to give you an analogy and you tell me if it captures it. Imagine you're baking a cake — you read the recipe, you put everything in the oven, but you're not allowed to touch it, open it, or check anything inside until it comes out the other end. Whatever happens inside is a black box. Is that how materials get made today?
Jared O'Leary (06:53)
That's a really good way to summarize it. In battery electrode manufacturing, for example, the material goes through a dryer that looks like an oven, and there is no visibility into what's happening while it's in there. That is the exact analogy. In the case of turbine blades, the turbines are coated and then go through a furnace-like structure for testing. Now, it's not quite a completely closed oven — there are cameras and things that exist today. But if you're baking a cake and the outside looks fine, you won't see that the inside has a catastrophic issue. Or maybe you will see the outside start to deteriorate — but only because there was a problem inside five minutes ago that you couldn't see, and now the entire batch is ruined.
Bogdan Cristei (08:05)
I see. And I assume there's a cost to that. What does this blindness actually cost — not in abstract terms, but give me a concrete example.
Jared O'Leary (08:16)
The most straightforward example is yield loss. If my part becomes bad at step one, and I don't catch that until step twenty, I lose all the money I put into processing it — the material added, all the extra value from steps two through nineteen. But I think it's easy to improperly categorize this as purely a yield problem. There are other aspects that provide even more value. Not all uncaught errors have the same cost. There's a difference between shipping a low-performing battery where you have an unhappy customer who can't charge as fast as they want, versus shipping a battery that catches fire and you have to recall thousands of cars — and lose billions of dollars. There are many well-documented cases of that.
But outside of yield and safety, there's another thing people underestimate: manufacturers know that certain defects and variations will come through their process that they have no visibility into and no control over. So they design the end product to be more robust to that — and there's a cost to that. For example, people make battery electrodes about 30% thicker than ideal performance would require, because they need that extra thickness to be robust against certain defects. That's a yield loss in its own right.
And then there's something that I think is really understated: having more information about the product you ship can allow it to operate better. Jet engines, for example, are run at temperatures above the melting point of their components. To prevent that, manufacturers coat their turbine blades with thermal barrier coatings — a complex multi-step procedure involving surface preparation, coating, growing a layer on top of that coating, and then a third layer. There's no good way to measure the quality of those coatings without destroying the blade. So manufacturers destroy about 1% of blades and then just assume some safety margin across the rest. But if they simply knew the uniformity of every blade, they could place them in positions where they'd run the engine hotter and more efficiently. We were told that with that information alone, they could run engines up to five percentage points more efficiently.
The same logic extends to batteries — if you had more information, you could create custom charging and discharging protocols for each battery rather than averaging across all of them. People really underestimate how much extra information about a product can improve its end operation. And finally, obviously, if you have more information, you can correct and optimize your manufacturing processes in real time. If you see drift, you can correct it. If you see something about to drift, you can get ahead of it.
Bogdan Cristei (13:53)
When I hear turbines, batteries, semiconductors — these aren't new industries. People have been working on them for decades. So why is this a 2026 problem and not a 1991 problem?
Jared O'Leary (14:08)
Versions of these problems did exist in 1991. But now we live in a world where performance requirements are much bigger. The most powerful computers that filled a room in 1990 are less powerful than an Apple Watch today. That requires a completely different level of complexity. We also have a world that wants to move away from fossil fuels, so we have to adapt accordingly. The short answer is that we've gotten very good at making more complex materials, but the blind spot has grown with that complexity.
Bogdan Cristei (15:45)
These problems aren't new. Smart people have been working on measurement and inspection for decades. Why hasn't this been solved?
Jared O'Leary (16:05)
People can see more inline than they could before — they just still can't see very much inside the manufacturing line. Measurement to date has always followed a philosophy of: I'm going to have a machine interact with a sample and elicit a very specific signal, in a very specific circumstance, to measure a very specific property. These tools were never designed to be wholly generalizable. Traditionally they've relied on cameras, x-rays, ultrasound, and lasers — all of which are phenomenally more advanced than they were thirty years ago. But even with those advances, there are many properties of advanced materials that these tools simply were not designed to measure. A camera was never designed to measure the conductivity of a material. An x-ray was not designed to measure chemical composition. If I have a product that's already made and I'm willing to destroy it and don't care how long it takes, I can measure almost anything I want. But for inline, non-destructive measurement — yes, there is definitely a physics gap.
Bogdan Cristei (18:27)
So when we talk about physics, we have to talk about cold atmospheric plasma. Can you explain what that actually is as if I know nothing about it?
Jared O'Leary (18:42)
Let's start with plasma in general. Plasma is occasionally considered the fourth state of matter — it's the most common state of matter in the universe, something like 98% of it. Plasma is basically a charged gas. The sun is a plasma. All the stars are plasmas. When we say cold atmospheric plasma, "cold" means a little above room temperature, and "atmospheric pressure" means it can exist in open air — not in space. An example of cold plasma that people encounter every day without realizing it is neon signs. Neon signs are atmospheric-pressure charged versions of neon gas.
Bogdan Cristei (19:50)
Interesting. Walk me through how you're using this — maybe as a kind of fifth sense — to get information about the world. Talk to me about the sensing technology you've been developing.
Jared O'Leary (20:02)
Our core product is fundamentally a measurement tool, with software alongside it. Its core mechanism is that we expose material samples to cold atmospheric plasma. These plasmas are unique in that when they come into contact with any material, they induce synergistic chemical, electrical, and thermal plasma-material interactions. So you have this scenario where the plasma is hitting a material, the material responds to the plasma, and the plasma responds to the material. The entire underlying thesis of this company is that those interactions carry a lot of structural and chemical information about the material. And if we can record those interactions sufficiently well with specialized detectors, we can translate that information into multiple critical material properties, performance predictions, quality control classifications — basically any piece of information that's relevant to the manufacturer.
Bogdan Cristei (21:50)
Interesting. So the interaction itself generates data that you could think of as a multi-dimensional signal that encodes the material's fingerprint.
Jared O'Leary (21:56)
That's a really good way to put it. We're calculating hundreds of thousands of individual signals per millisecond, and the combination of those signals really does represent a unique material fingerprint.
Bogdan Cristei (22:12)
So then you have a lot of data, you do machine learning on it, and you can extract useful things from it.
Jared O'Leary (22:24)
Yes — and this is where I want to be precise. It's very easy to hear "there's a lot of data" and assume you just throw a neural network at it. I'm not interested in that approach. The information we capture is physical. The goal of what we do — and we do use machine learning tools — is really to discover what's happening physically within the system. It's physics-informed machine learning, not just pattern matching.
Bogdan Cristei (22:53)
What does the signal actually tell you? What can you measure that you couldn't measure before?
Jared O'Leary (22:57)
In the context of battery manufacturing, we can measure things inside the manufacturing line in real time that previously couldn't be measured at all — simultaneous measurements of porosity, adhesion, chemical composition, and conductivity. We can identify the presence of sub-30-micron metallic contaminants in the bulk of electrode materials. We can find spatial defects on the edge of electrode materials and pinhole defects within them. Right now, the only thing most lines can measure inline is the thickness of the material — and only for single-sided electrodes. The entire industry is moving toward double-sided, which is a different challenge entirely.
Bogdan Cristei (23:52)
I want to ask about the phrase "manufacturing intelligence," because I think people hearing this for the first time might think you're just describing a better measurement device. What's the difference between a smart sensor and a manufacturing intelligence platform?
Jared O'Leary (24:10)
Let's take a well-known sensor — an ellipsometer, which measures the thickness of materials and was very commonly used in semiconductors. The fundamental output is the thickness of the material. You know how thick it is, and that's what you know. There's nothing else you can do with it later. For us, the first-level output is this material fingerprint. At that first level, yes, we can get properties like thickness, density, conductivity, and so on — and that's valuable in the moment. But only part of the fingerprint is used to derive those properties. What's interesting is that you can store that fingerprint, and six months later correlate it to the downstream performance of the end product in the field. After six months, you learn something new about your manufacturing process that you didn't know before. The same raw data can be used at one level for property prediction, at a second level for broader process intelligence — what's new about my product, process, or equipment — and at a fourth layer for actually optimizing and changing the manufacturing process in real time.
Bogdan Cristei (25:58)
Got it. So instead of setting a recipe, running it, and looking at the end result, you now have a continuous understanding of what's happening throughout the process.
Jared O'Leary (26:08)
Exactly. And it's really meant to increase the manufacturer's understanding of their processes over time. People think: if I build a factory and it's running well, I can just copy it to a new location. But that copying process is actually really complicated. Even companies like TSMC famously take two years to copy a factory, because so much information about how the process is run is legacy knowledge that isn't well documented. Anyone who works in advanced manufacturing can tell you some story about how a seemingly arbitrary change in their process made a huge difference. Not because the process is magic, but because there are many things that aren't understood. The more you can give manufacturers visibility into their processes, the faster they can copy them, the more money they can make, the faster they can go to market, and the more robust they can be against unexpected changes.
Bogdan Cristei (27:28)
So it's fragmented. Each engineering team has their local optimizations, their process, their control values — "don't touch it and it works."
Jared O'Leary (27:34)
Right, but there are also little micro-changes happening constantly. And a lot of times those are made by the technician who isn't necessarily communicating them to the engineer. There are stories of technicians at big battery manufacturers literally sticking their fingers into the slurry used to make electrodes and tasting it — and based on that taste, making certain adjustments. They probably can't fully articulate why the taste correlates to better output, but they know it does. People imagine manufacturing parameters are locked in and never change. That's not true. There are small adjustments every day, and they're not always communicated. A lot of incomplete information exists at every level.
Bogdan Cristei (29:30)
Back to AI for a moment — there's a lot of hype right now around digital twins, predictive maintenance, autonomous factories. Where does SirenOpt fit in that picture? Are you competing with that wave or enabling it?
Jared O'Leary (29:49)
We're enabling it. Let's be precise about what a digital twin is. People think it's literally a digital copy of a factory, but it's really a model — one that has inputs and outputs that mirror or approximate what's going on in the real world. Those models are better with more data and more visibility. We certainly enable that. But there's a big difference between what we do and what most digital twin approaches do. A lot of digital twin implementations take telemetry data — camera footage, environmental readings, location data — and try to correlate it to some output in a purely data-driven way. What's fundamentally different about us is that the data we collect is inherently physics-grounded. Because it's physics-grounded, it's meant to extrapolate and be interpretable beyond the immediate data. And when people talk about autonomous factories, they imagine some universal AI that makes every factory autonomous. But anyone who works in these environments knows these models are best when tailored to specific scenarios. In order to really use our data well, you have to understand it deeply. So we also plan to develop real-time process control software — and we're uniquely suited to do that given our domain knowledge of this raw data.
Bogdan Cristei (32:14)
Let's talk about who's actually using this today. Some of your customers are serious Tier-1 manufacturers — these are not companies that run experiments for fun. What does deployment actually mean?
Jared O'Leary (32:29)
The only customer we can mention publicly right now is one that has invested in us — Jaguar Land Rover, which is a subsidiary of Tata. But what I'll say is we have customers who are genuine Tier-1 manufacturers. The internal champions at these companies — in one case someone who has led non-destructive testing for 20 years — are people who have seen every tool that exists. And there is a novelty aspect to what we do: we are the first company to ever use cold plasma for measurement. Customers rightly question that. We always say yes to validation. The first way we interact with customers is we measure their samples and send them back for verification. Sometimes customers will spend four months verifying those measurements with six different instruments, measuring each material multiple times. But once they have that confidence, they commit seriously. Our tools are currently running at multiple Tier-1 manufacturing sites in North America, Europe, and Asia.
Bogdan Cristei (34:43)
So they send you a sample, you run it through your process and give them a dataset, they take it back and verify with their own instruments — something you can do in a few hours takes them months to confirm.
Jared O'Leary (35:08)
Yes. And sometimes there's value in the verification process itself. Customers who dig through our raw data often discover things we weren't even measuring for. One example: we ran a proof of concept focused on measuring certain properties of a separator coating. While they were validating those results, the customer's team looked through our raw data signals and realized we could actually determine the anisotropy of pores in their separator — a significant problem for them that they had never even thought to ask us about, because they didn't believe any tool could capture it.
Bogdan Cristei (36:42)
So how do you make money? On the surface this sounds like a hardware company, but I suspect the story is more interesting than that.
Jared O'Leary (36:45)
We sell the hardware — a capital equipment acquisition fee, in line with how all advanced manufacturing equipment is sold. The industry standard is a hardware sale plus a ten-year annual support fee over the lifetime of the tool. On top of that, we sell software. The core modules are: the operating system used to run the tool and generate model predictions; an analytics suite for looking through raw data, finding patterns, and correlating to operational parameters; and a process control module — the decision-making layer — because there's a fundamental leap between "here are the correlations between these inputs and outputs" and "here is what we recommend you change, or let us change it automatically."
Bogdan Cristei (38:21)
Walk me through the economics at a high level. If I'm a turbine manufacturer and I write you a check, what actually happens to my P&L?
Jared O'Leary (38:31)
In turbines specifically, the most immediate value is this: turbine manufacturing is highly regulated, so you legally cannot ship poorly coated components. Right now, you coat a blade, test it at the end, find it isn't good, and then strip it back to bare substrate and recoat — that's an expensive rework cycle. With our system, you can test each layer as it's applied, so your throughput improves and you catch catastrophic errors before they compound. That gives you yield improvement and throughput improvement together. In the medium term, you have more information about your turbines so you can operate them more effectively. And in the longer term, you have an entirely new class of data to describe your manufacturing process, which opens the door to truly transformational process changes.
In terms of ROI, you should think about yield and throughput improvement, reduction in destructive testing — you destroy fewer samples for validation — improved performance of the end product, meaning you can sell it for more or capture more market share, and an additional safeguard against catastrophic failures. These manufacturers pay hundreds of millions in insurance payouts and liabilities due to undetected defects. That exposure shrinks significantly with real inline measurement.
Bogdan Cristei (41:04)
One comparison I keep hearing when people mention SirenOpt is KLA in semiconductors — a company now worth over $200 billion that essentially owns the measurement intelligence layer for chip manufacturing. Is that the right analogy?
Jared O'Leary (41:31)
I think the comparison to KLA is fair at the initial property prediction level. KLA does classic metrology measurements and they do them very well. But their tools don't inherently lend themselves to additional layers of value through analytics and process control software — which is why KLA doesn't sell meaningful software, and neither do Keyence, Applied Materials, or most incumbents. The more appropriate analogy in a lot of ways is actually the manufacturing intelligence companies that have built new observability layers for physical systems — companies that collect a novel mode of data and use it as a raw data layer that enables a range of downstream applications.
Bogdan Cristei (42:43)
This seems like an obvious opportunity. Why isn't a well-funded incumbent building this? What makes SirenOpt unique?
Jared O'Leary (43:01)
People are certainly trying to build better measurement tools. But our uniquely differentiating aspect is cold plasma itself. It enables multimodal, inline, non-destructive measurement in a way that creates a physics gap that the world has never seen closed before. The fair question is: why doesn't anyone else use cold plasma for measurement? The answer has a few parts. First, I co-founded this company with my former PhD advisor based on research at UC Berkeley. We weren't originally trying to make a cold plasma sensor — we were trying to control cold plasma for biomedical applications. We were the first lab to ever apply machine learning to cold plasma in any capacity, and the first to try to control it. That control is critical because it allows for a consistent measurement — and nobody had tried to do that seriously, partly because the machine learning techniques required only became mature enough relatively recently. Second, translating plasma-material interactions into useful information also requires physics-informed machine learning techniques that weren't available until recently. And third, even if someone wanted to replicate what we've built today, there is a massive time cost involved in collecting the proprietary data and calibration models that drive our tool's performance. You have to measure a lot of materials across a lot of different companies and environments to really learn how your tool behaves. That takes years and can't be shortcut.
Bogdan Cristei (45:16)
You and Ali created this at UC Berkeley. At what point did you look at the research and say — this could be a company, not just a paper?
Jared O'Leary (45:26)
The full story is that we were developing these cold plasma systems for a completely different application. Over the course of that research, we essentially accidentally realized the system could be repurposed as a measurement tool. But the moment we saw that, we immediately knew it could be a company — because our other research had made us intimately familiar with advanced materials manufacturing and how much limited visibility into production lines affects how those lines are controlled. There was never a moment where we thought this would just be a paper. Inventing a new measurement product isn't a research paper. It's a company.
Bogdan Cristei (46:14)
Where can people learn more about SirenOpt?
Jared O'Leary (46:18)
They can go to our website at www.sirenopt.com. We also just gave a talk at the Volta Foundation presenting results collected with a collaborator showing that we can measure critical material properties of various battery electrodes.
Bogdan Cristei (46:36)
I'll put that link in the show notes. Awesome — really good to have you on. Thank you so much for your time today.
Jared O'Leary (46:38)
Thanks very much.