Core insights building a Vertical AI factory

Plastics is a ~$600B global market where material decisions still rely on fragmented PDFs, tribal knowledge, and disconnected tools. Engineers are expected to move faster, reduce risk, and meet rising cost and sustainability constraints, yet the underlying systems for researching, specifying, and sourcing materials haven’t meaningfully evolved in decades.

Over the last year, we’ve been building Herman AI at Plastics.com, a vertical AI sourcing platform purpose-built for plastics and materials engineering. Today, Herman can reference over 46,000 plastics, over 100,000 pages of expert engineering knowledge, and 10,000 categorized plastics industry websites, enabling engineers to research, evaluate, and order materials in a single workflow. Since last October, Herman’s usage has grown ~40% month-over-month to >300 companies, helping engineers find materials 3× faster and enabling teams to receive quotes from 100+ suppliers within 24 hours.

Along the way, we’ve gained hard-earned insights into the limits of LLMs while solving those challenges with AI agents embedded in real engineering and procurement workflows.

First: The GPT-style user experience isn’t enough.
A simple chat interface breaks down quickly in B2B engineering environments. Engineers and procurement teams need collaboration, shared analysis, versioned decisions, integrations with specialized tools, and support for real workflows—not isolated prompts and answers.

Our answer was to embed Herman into a custom-built marketplace that seamlessly connects engineers to suppliers on Plastics.com.

Second: Engineers need workflow completion, not isolated tasks

Most frontier AI tools today optimize for individual moments: answer a question, or summarize a document. That’s useful—but it’s not how engineering work actually happens.

 AI for engineering must be designed around end-to-end workflow completion with high trust, not prompt-level intelligence. Memory, project structure, collaboration, and system integrations matter as much as model quality. In practice, this means building AI that behaves less like a chatbot and more like a continuously running process that is embedded across engineering and sourcing workflows.

Third: an ontology is necessary, but nowhere near sufficient.
Domain structure matters. You need a shared language for materials, properties, processes, suppliers, test methods, and trade-offs. But an ontology alone doesn’t tell you what matters, what’s credible, or what to recommend in a specific engineering context.

That’s where three things became critical for us:

1. Material Rank
Engineers don’t just need results; they need prioritized results. Material Rank encodes relevance, applicability, supplier credibility, performance fit, and real-world usage signals. Without ranking, AI outputs feel impressive but unusable. With it, engineers move faster and trust the system.

2. Engineering Knowledge (not just data)
Datasheets are incomplete, inconsistent, and often contradictory. Engineering knowledge lives in the gaps: how properties interact, what fails in production, which substitutions actually work, and where risk hides. Capturing and operationalizing that requires more than retrieval—it requires reasoning grounded in domain context.

3. Engineering Memory
Engineering work is cumulative. Decisions depend on past constraints, prior failures, approved materials, internal standards, and supplier relationships. We’re building engineering memory alongside our ontology so Herman can remember why a decision was made—not just what was selected—while keeping that memory private and customer-controlled.

The bigger realization: Vertical AI isn’t a prompt. It’s an AI factory.

To make this work, we’ve had to build infrastructure that spans:

·       query decomposition, intent identification, and orchestration

·       multimodal domain knowledge processing and encoding pipelines

·       ontology enrichment using physics-based and genetically optimized ML

·       multimodal graph retrieval of domain knowledge

·       private, contextual memory

·       engineering-specific evaluation frameworks

·       and custom RL models embedded directly into workflows

Turning natural language into reliable, repeatable engineering outcomes requires policies, guardrails, evaluations, and constant feedback loops. Agentic AI only works when the system knows when not to answer, when to ask clarifying questions, and how to reason inside domain constraints.

How are we doing:

Plastics.com delivers value to our customers by directly integrating Herman AI into their engineering and procurement workflows. Customers trust our verticalized agentic AI platform because we outperform frontier models by around 48% across expert-graded and evaluated NCEES plastics-related engineering questions, has 20% lower hallucination rates, and are 12.8x faster when answering engineering-related questions.

The payoff:
This approach lets us rapidly innovate, increase data leverage, and refine our strategies as new foundation models emerge, without rewriting the product, which continuously improves accuracy, trust, and speed for engineers sourcing real materials under real constraints.

Where are we going:

We’re focused on reducing friction in agent-based transactions and logistics by using AI to deeply integrate existing engineering, procurement, and supplier workflows. We’re also excited to be launching agentic simulation, cost analysis, and CAD tools purpose-built to make designing complex plastic parts easier for the 45M CAD users worldwide. Our team is striving to make the world a better place through greater education about sustainable options, bringing the tool to other verticals.

Vertical AI is harder than it looks.
But when it’s done right, it stops being a demo—and starts becoming infrastructure.

We built Herman AI for engineers → Plastics.com for procurement → Workflows to become infrastructure.

Author: Dale Thomas, CEO