The Research Wasn't the Product. The Judgment Was.

The too long, didn’t read: I set out to build a research system and ended up with a new product and a new company. The process of building it turned out to be more interesting than the research it produced.

What Layer2C is

Layer2C is a set of websites built around two models I created over the past year. The first is the Fourth Cloud readiness assessment, which is essentially private cloud for AI. The second is the 4+1 AI infrastructure framework, which is what this post is about.

The 4+1 model is an eight-layer stack. It lets an organization break down its AI estate from storage, compute, and networking, up through model selection and management, through the agentic layer, all the way to the application and the business value it’s supposed to deliver. The intent was to give enterprise architects (EAs) a common language that works across any AI vendor. An EA can look at a vendor and quickly place it in a stack they already understand, and just as quickly see where their own internal systems plug into that same stack.

From framework to instrument

Once I had a model that could describe any AI vendor, the next step was obvious. Build an assessment instrument. I’d already published an RFP template based on the 4+1 model, and the honest critique of that template was that it didn’t score anything. I’ve since fixed that, because the research system forced me to build the scoring.

Then came the question every analyst eventually asks. What are my opinions? If I applied this instrument to the obvious cohort of AI infrastructure companies, how would they compare? So I started with Dell, NVIDIA, HPE, and the other core companies I already follow and cover.

Here’s how the system works. I built an agentic workflow using Claude Code. Claude goes out and researches a vendor through its public documentation, compares what it finds to the 4+1 model, which is an open model with published definitions, and rates the vendor at each layer. Then I go back through the result. I add context, I correct errors, or I write a rule about how to deconflict the definition with the reality of the market.

That last part is the important part. I was codifying my judgment into the workflow. And it wasn’t just guardrails in a prompt. I’d already built the deterministic-code-in-the-loop concept, so this was prompt guardrails plus code that could be applied and enforced, every run, without me watching.

Scale breaks everything, until it doesn’t

Over the following weeks and months, I expanded the set of vendors. Not just the number but the diversity. We started with traditional infrastructure OEMs and infrastructure-as-a-service providers, then moved to software companies, AI labs, neoclouds, and streaming data platforms. Anyone with an AI story, from a product marketing perspective, could now be assessed against the same instrument.

As you’d expect, the further in we went, the more nuance surfaced. Each new category of vendor exposed a place where the definitions needed my expertise to deconflict, refine, or extend the instrument. And each time I did that, the landscape of vendors the system could handle got wider. Today we track 46 vendors with this instrument.

That was my aha moment. The product isn’t the 46 assessments. The research is interesting, and I’m proud of it. However, the thing with real value is the ability to codify expert judgment in a way that scales, because as I’ve said repeatedly, scale breaks everything. This was the first time I’d built something where my judgment scaled without breaking.

Does it scale past me?

The obvious follow-up: does it scale beyond one person? The folks at Kamiwaza sponsored Lab 015 in the Layer2C Labs series, and the question I proposed was exactly that. Can I hire a second analyst, extend my judgment to them, and have them run the system? Disclosure: Kamiwaza paid for that lab. The conclusion is mine.

Too long, didn’t read: yes, it scales.

Knerd AI

The formal product is a sister company called Common Knerd, and you’ll find it at knerd.ai. It’s the formalized system for codifying expert judgment across an organization. And yes, I’m a founder, so weigh what follows accordingly.

Let me abstract this beyond research. Look at the management consulting governance model: partner, managing director, director, senior manager, senior associate, associate, junior associate. Consulting firms became experts at taking a framework the partner created and applying it consistently across every one of those levels. That conformity is what ensures consistency across people, projects, and clients. We took that model and applied it to an agentic platform. Capture the expert-level judgment, codify it, and iterate on it over time, exactly the way I did with Layer2C.

Here’s a second proof point outside the vendor assessments. I run something called the design challenge. I take a vendor’s marketing material and challenge them: I’ll produce a design an EA can look at with reasonable comfort that it’s applicable to their environment. Doing that well takes expert knowledge of lab work, of vendor marketing, of the actual technology, and of the interview questions you have to ask the customer first. It’s a CTO Advisor shaped problem.

So what happens when I have a seller who isn’t versed in all of those areas, but I need them reaching as many buyers as possible? That’s Knerd AI applied to a sales motion. There’s an agentic workflow with that skill codified into it, and a sales representative can run it without me validating every output. Sure, I might look at the resulting design and tweak it, or ask the rep to tweak it because something is unrealistic. But that becomes a continuous improvement loop, and it lets me scale the one area that has frustrated me most over the past few years: sales, when most sellers don’t have my ability to architect a solution for the customer.

This isn’t vaporware. We built the system, we use the system, and the 46 vendor assessments on Layer2C are its output. If you want to see the process, there’s an interactive walkthrough at knerd.ai/demo. If you want to kick the tires on the actual prototype, contact me and I’ll walk you through it. I’d love your feedback.