My podcast guest this week is Dr. Maha Achour, CEO and Founder of Kodamai. Maha and I chat about the shifting landscape of AI and how we can move beyond cloud and centralized data centers to a more distributed, agentic future. We discuss why governance must ‘travel’ with AI agents, what it means for trust to be architectural rather than procedural, and how Kodamai’s ‘math-first’ approach is helping engineers build verified, auditable AI systems that can thrive in the real world.

Links for July 31, 2026

More information about Kodamai

Connect with Kodamai via LinkedIn 

Connect with Maha Achour via LinkedIn 

Click here to check out the Fish Fry Archive.

Click here to subscribe to Fish Fry via Podbean

Click here to get the Fish Fry RSS Feed

Click here to subscribe to Fish Fry via Apple Podcasts

Click here to subscribe to Fish Fry via Spotify

Amelia’s Weekly Fish Fry  – Episode 692

Release Date: July 31, 2026

Hello everyone, and welcome to Episode 692 of Amelia’s Weekly Fish Fry, brought to you by EEJournal.com and written, produced, and hosted by yours truly, Amelia Dalton.

This week I’m joined by Dr. Maha Achour, CEO and Founder of Kodamai.

Maha and I are diving into the shifting landscape of artificial intelligence, moving beyond the cloud and centralized data centers toward a more distributed, agentic future. We discuss why governance must travel with AI agents, what it means for trust to be architectural rather than procedural, and how Kodamai’s math-first approach is helping engineers build verified AI systems that can thrive in the real world.

So, without further ado, please welcome Dr. Maha Achour to Fish Fry.

Amelia: Hi, Maha. Thank you so much for joining me.

Dr. Maha Achour: Thank you for having me.

Amelia: Absolutely.

For years, we’ve thought about AI as something that lives in the cloud or inside a centralized data center. But now AI is becoming increasingly distributed. What’s driving this shift, and how does it change the way we design intelligent systems?

Dr. Maha Achour: That’s a great question. I like to answer it in three parts.

First, let’s look at how we got here.

If you look back through history, you can think about Socrates and the Socratic method. At that time, there were no machines or computers—just the human brain. Socrates would gather groups of people and encourage conversation, asking questions and challenging assumptions. Through that argumentative dialogue, he stimulated critical thinking. In many ways, that’s an example of distributed intelligence using only human minds.

I think it’s important to remember that the human factor will always remain incredibly powerful throughout this journey.

As a particle theoretical physicist, I’ve always appreciated how pure mathematics has enabled us to understand both the microscopic and macroscopic worlds decades before experiments could verify those discoveries. That’s why I wanted to learn that language during my Ph.D. at MIT.

Interestingly, my first publication in 1989, which used mathematical methods to understand the physical networks of neurons in the brain, was later featured on the cover of Science magazine. It reinforced for me that elegant, simple mathematics—combined with human intelligence—can play a major role in understanding intelligence itself.

Over time, thanks to the internet and the enormous amount of distributed compute power available in data centers and servers, humans and industries generated massive amounts of data. Large language models were trained on that data and initially operated almost entirely in centralized environments.

As AI evolved, intelligence naturally became more distributed because it needed to be accessible and processed in real time. That’s where edge computing and client-side processing became increasingly important.

What’s truly disruptive today is the rise of agentic AI.

We’re moving beyond simply distributing data and models. We’re now distributing processes—whether they’re enterprise workflows, robotics, or autonomous vehicle systems. Processes are becoming the new unit of intelligence, and that makes agentic orchestration at scale critically important.

This is where the industry is heading. To deploy these systems successfully, governance and correctness have to be built into the architecture itself.

Amelia: As AI agents move into the field and begin making decisions closer to where data is generated, governance becomes much more complicated.

Why is it important that governance travels with the agent rather than remaining in a centralized location?

Dr. Maha Achour: Excellent question.

I could answer this by diving into the mathematics behind Kodamai, but I think an example from the semiconductor industry is much easier to understand.

I’ve designed RF chips myself, so I’m very familiar with that development process.

Think about today’s semiconductor design flow. You have an incredibly complex and expensive process involving design, verification, testing, compliance, intellectual property protection, governance, quality assurance, and auditing before a chip is finally manufactured.

Now imagine a future where, instead of building chips layer by layer on wafers, we use nanoscale composable building blocks.

Each of these building blocks would have intelligent interfaces. They would know how to position themselves and compose the final chip based on clearly defined engineering requirements.

In that world, governance, compliance, and auditing wouldn’t happen after the design was complete—they would be embedded directly into the building blocks themselves.

That’s exactly the model we’re pursuing for software-based agentic orchestration.

Today, organizations build AI agents, test them, verify them, and then hope they don’t make mistakes.

Our vision is fundamentally different. Governance, trust, and auditability become part of the fabric of the orchestration system itself.

Amelia: We also need to talk about trust.

We’re hearing more people say that trust must be architectural rather than procedural. What does that mean in practice, and how can engineers build trust directly into AI systems from the ground up?

Dr. Maha Achour: That’s another excellent question, and it builds directly on the semiconductor analogy.

Today’s enterprise workforce follows well-defined procedures and strives to operate in a trustworthy, coordinated way. AI agents need to do the same.

The challenge is that today’s agents are built primarily from code, tools, and machine learning models. They take data, combine it with tools and workflows, and produce agents.

But there’s a much larger transition happening underneath all of this.

We’re moving away from decades-old, data-centric architectures toward systems that organize entire enterprise processes.

The problem isn’t simply building LLM agents or reinforcement learning agents. The real challenge is scaling them.

How do you enable thousands of agents to collaborate, make decisions at the edge, and propagate throughout an organization while maintaining correctness?

To accomplish that, you need an underlying agentic fabric that transforms business processes directly into software—not simply collections of functions and classes.

That’s exactly what we’ve built with the Kodamai Calvin Grove platform.

At its core, Calvin Grove transforms business engineering into software engineering.

Workflows become first-class citizens. They’re composable, much like the nanoscale building blocks I described earlier.

These composable modules communicate through typed interfaces, which we call contracts.

Types are familiar to software engineers at the implementation level, but in formal computer science they have much broader significance. By leveraging typed interfaces and mathematical contracts, we eliminate entire classes of runtime errors before execution ever begins.

That’s what we mean by correctness by construction.

Instead of adding trust after a system has been built, trust becomes an inherent property of the system itself.

Today, much of our hardware infrastructure still reflects the Industrial Revolution—a world where machines replaced human muscle.

We’re now entering what I believe is the Intelligence Revolution, where software augments human intelligence across every layer of computing.

To fully realize that future, both our software and our underlying infrastructure need to evolve together.

That’s the new paradigm.

It’s the paradigm that will ultimately drive the evolution from today’s AI toward AGI and, eventually, ASI.

I’m very excited that Kodamai and the Calvin Grove language are at the forefront of that transformation, and I hope we’ll be able to make a meaningful impact on both software today and hardware in the future.

Amelia Dalton: Looking to the future, how do you see distributed AI, embedded governance, and architectural trust reshaping the relationship between hardware, software, and autonomous agents? And what should engineers be doing today to prepare for that future?

Dr. Maha Achour: That’s an excellent question.

I was actually at UC San Diego a few weeks ago, and the moderator asked me this very same question because the engineers in the audience were eager to understand where all of this is headed.

It really goes back to something I mentioned earlier: we need a shift in engineering education. We have to educate the next generation of engineers not only on the hardware and systems engineering practices we’ve inherited over the past several decades, but also empower them to think beyond those traditional models.

Engineers need to understand how this AI revolution in software will influence everything from materials science and semiconductor manufacturing to governance, system design, and product development.

If we think about today’s tiered computing architecture—the core, the edge, and the client—we also have to ask how governance and trust are propagated across those layers.

A great example is the autonomous vehicle.

An autonomous vehicle is essentially a robotic system that operates with very little human intervention. It has to make decisions safely whether it’s operating within the vehicle itself, at the edge, or while communicating with cloud infrastructure.

Traditionally, we achieved safety by collecting enormous amounts of data, training sensor fusion systems, and trying to account for every possible corner case.

Now we’re entering a different era.

We’re embedding generative AI, reinforcement learning, and agentic orchestration directly into the vehicle so it can maintain safety and governance continuously—even when connectivity to infrastructure or cloud services is unavailable.

Once vehicles begin communicating through V2V (vehicle-to-vehicle) and V2X (vehicle-to-everything) networks, the challenge becomes even greater. Governance and trust must be enforced across an increasingly distributed infrastructure.

That’s why an underlying orchestration fabric is so important. It allows agents operating across every tier to coordinate safely and consistently.

In other words, you don’t impose governance after something goes wrong—you build governance directly into the system from the very beginning.

Ultimately, that architectural fabric must exist in both software and hardware. That’s exactly what Kodamai’s Calvin Grove compositional modules and typed interfaces are designed to enable.

Amelia Dalton: That makes sense.

Given Kodamai’s focus on AI agent platforms, how does your approach specifically address embedded governance and architectural trust—particularly when agents are operating in distributed edge environments?

Dr. Maha Achour: Great question.

I’ll introduce a few technical concepts that I think your audience will appreciate.

The Calvin Grove language and orchestration platform are built on a distributed, typed microservices framework.

Those three words—distributed, typed, and microservices—carry a lot of meaning.

Within Calvin Grove, a component we call the daemon acts as a hypervisor for entire systems of AI agents.

Above the individual agents sits another class of agents that we call meta-agents. These privileged agents oversee managers, executors, collaborators, decision-makers, and other specialized agents throughout the system.

The daemon coordinates all of them.

Even when agents are operating remotely or with intermittent connectivity, every component is implemented within the same language and governed by the same typed interfaces.

Those mathematical type contracts determine whether information can safely propagate throughout the system. If the interfaces don’t match, propagation is stopped before errors can spread.

Because of this typing discipline, Calvin Grove maintains correctness consistently across every computing tier.

Today, the cloud, the edge, and endpoint devices often function as largely independent systems.

Our approach replaces that separation with a unified architectural fabric that spans every layer.

The daemon enforces governance at every network boundary, allowing trust and correctness to become properties of the architecture itself rather than something added afterward.

Amelia Dalton: We’re also looking at an interesting intersection between academia, infrastructure, and private investment.

How can we bridge that gap and empower the next generation of engineers?

Dr. Maha Achour: The transition is going to happen gradually.

Universities are still largely teaching engineering through the lens of the Industrial Revolution. While curricula are evolving, there will inevitably be a transition period.

Industry has an important role to play.

Many of the companies leading AI development possess enormous resources and expertise. They should be partnering directly with universities and engaging with students much earlier.

Master’s theses, doctoral research, and university projects should be aligned with the challenges industry expects to face over the next decade, rather than focusing exclusively on topics that may become outdated by the time students graduate.

I believe universities, industry, and students need to work together over the next ten years to ensure this transition happens successfully.

Amelia Dalton: Let’s also talk about Kodamai’s math-first approach to agentic AI.

What does that look like in practice, especially when it comes to creating verified interfaces and auditable decisions in complex engineering systems?

Dr. Maha Achour: That’s probably a question my CTO or CISO could answer in much greater mathematical detail, but let me try to simplify it.

Most software engineers are familiar with types because every programming language asks you to define whether a variable is an integer, text, a floating-point number, or another data type.

Traditional programming languages type data.

When we talk about typed interfaces—or even typed agents—we’re talking about something fundamentally different.

These concepts come from a branch of formal computer science called category theory.

Category theory provides a mathematical framework for describing relationships, processes, and composition.

Think about following a recipe.

You have ingredients, a sequence of steps, and relationships between those steps.

Now imagine solving a physics problem.

Although the subject matter is entirely different, the abstract structure of the process can look remarkably similar.

Category theory focuses on those higher-level structures rather than the specific details of each discipline.

That’s the mathematical foundation behind Calvin Grove.

It combines category-theoretic formal methods with typed interfaces and type systems to create composable, verifiable software components.

At the same time, we augment those formal methods with machine learning and neuro-symbolic AI so we gain both mathematical correctness and intelligent reasoning.

Amelia Dalton: Fantastic.

Well, Maha, before I let you go, it’s time for your off-the-cuff question.

If you could have one meal right now—anywhere in the world, even if you needed a passport to get there—what would it be?

Dr. Maha Achour: Oh, wow!

I’d have to say Vietnamese cuisine. I especially love French-Vietnamese fusion. I couldn’t tell you the names of all the dishes, but that would definitely be my choice.

Amelia Dalton: Nice! That sounds wonderful.

Well, Maha, I think that’s all I have time for today. Thank you so much for joining me.

Dr. Maha Achour: Thank you.