In the rapidly evolving world of industrial AI, Kence Anderson and his team at Composable are on a mission to shift the paradigm from machine learning to machine teaching. On a recent episode of the Augmented Ops podcast, Anderson shared his insights on the limitations of traditional AI approaches in manufacturing and the potential for modular, explainable AI agents to transform operations.
Anderson’s journey in the field began with designing autonomous AI systems for various industries at Microsoft. “I personally designed over 200 of these kind of autonomous AI systems,” he said. “I’m talking about the control systems, the optimization systems. Closing the loop.”
However, he quickly realized that the predictions and analytics generated by machine learning models were only a small part of the decision-making process in industrial settings. What operators needed was a way to capture and transfer their expertise – the knowledge graph of skills practised to competence.
“I’m starting to use this phrase ‘engineered intelligence’ because to me, what engineers do is build things with building blocks,” Anderson explained. “They design complex systems from building blocks and I think that artificial intelligence should be that way too.”
That’s where Composable comes in. The company’s platform allows users to create AI agents by defining modular skills, goals, and orchestration through a drag-and-drop interface. Large language models bootstrap the process by enabling users to describe skills and goals in natural language.
The resulting agents have a three-layer architecture: a decision layer that makes choices, a perception layer that takes in sensory data, and an assistance layer that communicates with humans in natural language for explainability.
Anderson shared an example of how this approach was used to optimize the production of Cheetos. The extrusion process is surprisingly complex, requiring operators to control 25 different variables based on subtle cues like the moisture content of the corn. By breaking down the process into modular skills and sequencing them, Composable was able to train an AI agent that achieved expert-level performance in just two weeks – a process that normally takes human operators 10-12 years to master.
But the path to widespread adoption of AI in manufacturing is not without its challenges. Anderson emphasized the need for tech companies to approach operations with curiosity and respect, rather than condescension.
Engineers have a different mindset because of the responsibility. But to me, I find that exciting. So to me, there’s a whole new audience to quote-unquote evangelize or get to adopt technology that think very differently than software engineers.
Looking ahead, Anderson envisions Composable as an “operating system for skills” that enables an ecosystem of partners to solve problems across industries. By integrating with data providers, specialized algorithms, and technology partners, the platform could potentially transfer skills from one domain to another – like applying welding expertise from automotive manufacturing to construction.
It’s a bold vision, but one that Anderson believes is essential for unlocking the full potential of AI in manufacturing. “The next 18 months for us is about building the integrations in the ecosystem,” he said. With Composable leading the charge, the future of industrial AI looks more modular, explainable, and collaborative than ever before.
