Join our daily and weekly newsletter for the latest updates and exclusive content on industry-leading AI coverage. learn more
In the race to implement AI across business operations, many companies find that general-purpose models often struggle with special industrial tasks that require deep domain knowledge and continuous inference.
While fine tuning and search and augmented generation (RAG) can be useful, it is often not sufficient for complex use cases like supply chains. What Startup Articul8 is trying to solve is a challenge. Today, the company has debuted a series of domain-specific AI models for the manufacturing supply chain called the A8-Supplychain. The new model comes with the ModelMesh from ARTICUL8. This is an agent AI-driven dynamic orchestration layer that makes real-time decisions about the AI model to use for a particular task.
Articul8 claims that its model achieves 92% accuracy in industrial workflows and outperforms generic AI models in complex sequential inference tasks.
Articul8 started as an internal development team within Intel and was woven into an independent business in 2024. The technology came from Intel’s work, which built and deployed a multimodal AI model for clients, including the Boston Consulting Group, before ChatGPT was launched.
The company was built on a core philosophy that contradicts many of the current market approaches to enterprise AI.
“We build on the central belief that a single model will not be led by enterprise outcomes,” Articul8 CEO and Founder Arun Subramaniyan told VentureBeat in an exclusive interview. “We need domain-specific models to actually follow complex use cases in regulated industries such as aerospace, defense, manufacturing, semiconductors and supply chains.”
Supply Chain AI Challenge: When Sequences and Contexts Determine Success or Failure
Manufacturing and industrial supply chains present the unique AI challenges of general-purpose models struggling to handle effectively. These environments include complex multi-step processes where sequences between steps, branch logic, and interdependencies are mission-critical.
“In the supply chain world, the fundamental principle is that everything is a one-step pack,” explained Subramaniyan. “It’s all a bunch of related steps, and steps may have connections and recursions.”
For example, if a user is trying to assemble a jet engine, there are often multiple manuals. Each manual has steps that you need to follow in order, if not thousands. These documents are not just static information. Time-series data representing sequential processes that must be followed accurately. Subramaniyan argued that AI general models often fail to grasp these temporal relationships, even if they are enhanced by search techniques.
This type of complex inference tracks the steps to identify where the error occurred in the reverse direction – representing a fundamental challenge that a typical model is not simply built to handle.
ModelMesh: Not just another orchestrator, but also a dynamic intelligence layer
At the heart of Articul8’s technology is ModelMesh. This goes beyond the typical model orchestration framework to create what is called “agent agents” for industrial applications.
“ModelMesh is actually an intelligence layer that keeps determining, determining and evaluating things, just like one step at a time,” explained Subramaniyan. “It’s something we had to build completely from scratch because the tools out there aren’t really approaching doing what we have to do.
Unlike existing frameworks such as Langchain and LlamainDex, which provide predefined workflows, ModelMesh combines Bayesian systems with special language models to dynamically determine whether the output is correct, what actions to take next, and how to maintain consistency across complex industrial processes.
This architecture allows ARTICUL8 to describe as an industrial agent AI. This is a system that can be actively promoted not only in industrial processes.
Beyond Rag: An Overview Approach to Industrial Information
While many enterprise AI implementations rely on searched generation (RAG) to connect popular models to enterprise data, Articul8 takes a different approach to building domain expertise.
“We actually take the underlying data and break them down into components,” explained Subramaniyan. “Decompose PDFs into text, images, tables. For audio or video, we break it down into the underlying components and use different models to explain those elements.”
The company starts with Llama 3.2 as a foundation selected primarily for generous licensing, but transforms it through a sophisticated multi-stage process. This multilayered approach allows the model to develop a much richer understanding of industrial processes than simply obtaining chunks of relevant data.
The SupplyChain model goes through multiple refinement stages specifically designed for industrial contexts. For well-defined tasks, they use monitored tweaks. For more complex scenarios that require expertise, domain experts implement feedback loops in which responses are evaluated and corrections are provided.
How to use Articul8
It’s still too early for the new model, but the company has already claimed its partners with many customers, including IBase-T, Itochu Techno-Solutions Corporation, Accenture, Intel and more.
Like many organizations, Intel has launched the Gen AI Journey by evaluating general purpose models and exploring ways to support design and manufacturing operations.
“These models are impressive for open-ended tasks, but we discovered limitations as soon as they were applied to highly specialized semiconductor environments,” Srinivas Lingam, corporate vice president and general manager of Intel’s networks, Edge and AI Group, told VentureBeat. “They struggled with interpreting semiconductor-specific terms, understanding context from equipment logs, or inference from complex, multivariable downtime scenarios.”
Intel is deploying the Articul8 platform to build what Lingham called – Manufacturing Incident Assistant – an intelligent, natural language-based system that helps engineers and technicians diagnose and resolve downtime events on Intel’s Fab equipment. He explained that platform and domain-specific models ingest both historical and real-time manufacturing data, including structured logs, unstructured Wiki articles, and internal knowledge repositories. It helps Intel teams perform root cause analysis (RCA), recommend corrective actions, and automate some of the work order generation.
What does this mean for an enterprise AI strategy?
The ARTICUL8 approach challenges the assumption that a generic model using RAG is sufficient for all use cases for companies implementing AI in the manufacturing and industrial context. The performance gap between the specialized and general models suggests that technical decision makers need to consider domain-specific approaches for mission-critical applications where accuracy is paramount.
As AI moves from experiments to production in an industrial environment, this specialized approach may provide faster ROI for certain high-value use cases, but the general model continues to provide a wider range of specialized needs.
Source link
