The Audit Factory: How We Score 50 Personal Brands at Once Without Burning the Budget

Scripts pull the data. A cheap model drafts. Only the real judgment runs on the expensive model. Here’s the whole machine — so you can run it too.

By Dennis Yu · BlitzMetrics · Built with a Claude agent, June 2026

Here’s the problem we hit, and maybe you have too.

We score a lot of personal brands. A room of 50 dunkers at Dunk Camp. A table of contractors at DealCon. A workshop in Wichita. Every one of them needs the same thing — an honest read on whether Google can actually see who they are.

So we’d point our AI at the whole list and let it rip. And it worked. Until the meter ran out. Then you wait days for the cap to reset — or you pay hundreds in overage to keep going. I was kicking off jobs at 3am to catch the next window. At our pace that’s about $4,000 a month just to not wait.

The fix wasn’t a better model. It was realizing we were doing something dumb: running everything — the mechanical data-pulling, the boilerplate drafting, and the hard judgment calls — through the same premium engine. Like hiring a brain surgeon to take blood pressure.

Three tiers, three different workers

A personal-brand audit isn’t one job. It’s three, stacked — and they don’t need the same horsepower.

Underneath all three sits one plain folder of files — a Markdown-and-JSON “vault,” one folder per business. That’s the trick that makes the models swappable: the script, the cheap model, the frontier model, and a human in Obsidian all read and write the same files. You’re never locked to one vendor’s memory. Swap a model, keep the data.

The honesty rules live in the code, not the prompt

Here’s where most “AI scoring” quietly lies to you. You ask a model to score someone, and to be agreeable it hands back an 85. We don’t allow that.

That gate is a few lines of code that run on every audit — cheap model or frontier model, doesn’t matter. A sloppy model can’t hand you a dishonest 85 because the code clamps it to 79 and says why. Honesty you can prove beats honesty you hope for.

What the expensive tier is actually for: “earned but illegible”

This is the pattern we see over and over, and it’s the whole reason the judge tier exists. Someone has done the work — millions of followers, real skill, real reviews — but Google has no idea who they are. No entity home. No structured proof. The reach is real and the legibility is zero.

A cheap model can list the numbers. It takes a frontier model to look at 2.7 million followers next to a domain rating of 1 and say the true thing in a sentence the person actually feels. That judgment is worth paying for. The data-pulling next to it is not.

The cost surprise: it was never about tokens

When we did the math, the punchline flipped. The model API is the cheap part.

Read that again, because it’s the part people miss. A 50-business batch costs about three dollars in API either way. The reason it felt like $4,000 is that we were running it as interactive overage on a rate-limited plan. Move the bulk to a flat-rate cheap model and run the judge as metered API, and the same work runs for pocket change — and never makes you wait. The architecture buys you cap relief and unattended throughput, not token savings.

You don’t run this. You talk to it.

Nobody on our team should have to open a terminal. So the interface is a conversation. You tell the agent “run the audit factory on this list,” it runs the machine, and it hands you a visual dashboard — a stoplight wall of all 50, filter by red/amber/green, sorted worst-first. The Python underneath is just the engine the agent drives.

Same engine, three ways to drive it as you grow:

That third stage is how this becomes a team sport — how the young adults we train can each do what I do: score a batch, run a workshop, teach the next room. The engine doesn’t change. They just get a button and a login, and the secrets live in a vault instead of on anyone’s laptop.

Why we’re publishing the whole recipe

Because that’s the BlitzMetrics way. Everything we build, we document — so the next person can follow the example instead of guessing. This machine scores brands honestly, runs for the price of a coffee, and hands a beginner the same leverage I have. If you’re teaching people to grow with AI, that’s the shape of it: cheap where you can be, premium where it counts, honest in the code, and written down so it repeats.

See the personal-branding method View the audit scorecard wall

This article is itself an example of the rule it describes: an agent built the machine, and we documented it. Part of the BlitzMetrics meta series on how we work.