What is published here

Everything here comes back to one question: when an AI agent answers something, how much of what you paid for actually reached the answer? Performance, efficiency, cost and quality all fall out of that, and each one is treated as a measurement rather than a claim.

AEQ, the Agent Efficiency Quotient, is the instrument. It defines architectural efficiency as business value delivered per token consumed, and it is written as a specification so that someone who did not build it can run it and get the same number. The papers apply it and report what came back. Projects are the systems those numbers were taken from. Case studies are engagements where a build was measured against the path it replaced. Writing collects the shorter pieces published elsewhere.

A second thread runs underneath the cost work. Dave Blundin has been working on quantization since 1992, on the argument that compressing a model is what makes capable AI cheap enough to run anywhere. Treated as an engineering requirement rather than a talking point, that has to be tested on a real workload rather than a benchmark. The measured answer was not the expected one: a 4-bit quantized model passed rubric cells its own fp16 parent failed, and a smaller model certified more query classes than a larger one. Capability turned out to be a property of the model and the workload together, not of parameter count and not of precision.

The progression on the home page, silicon to software to platforms to agents, is a career rather than a slogan. Twenty years of it went into shipping hardware and software: chips through tape-out, Linux platforms owned end to end, and the program management that turns a specification into something a customer can buy. The last six were at Apple, running program management for the ChatGPT integration into iOS and the systemwide AI work around it. The research here holds to the same bar.

The work is Michael Valderrama's, and the resume has the specifics.

The method

Before a benchmark runs, the test plan is locked and published, so the numbers can't be tuned afterward. Then everything ships: the results, plus an errata log of mistakes caught along the way. Findings that cut against my own argument go up with the rest. The clearest case is in The Cost of a Question: below roughly 130 seats, per-seat SaaS beats the agent. That number hurts the pitch. It's in the paper anyway.