~/skunk $ ls ./posts

DISPATCHES FROM THE CHAOS

Hard-won lessons from the trenches — AI, automation, ops & ecommerce

~/skunk $ cat myspace-era-html.md

The Last Time I Wrote HTML, MySpace Was Winning

Every small-business site now looks identical: clean, professional, forgettable. Here's why I built this one scrappy instead — and why rough reads as honest.

2026-07-11METADESIGNAI

~/skunk $ cat ai-crew-two-weekends.md

Two Weekends, an AI Crew, and a Used Gaming Rig

How someone who hadn't written HTML since MySpace shipped a full Next.js site in two weekends — by running an AI crew with a used gaming rig as the labor.

2026-07-11METAAILOCAL-LLMDEVLOG

~/skunk $ cat the-privacy-wall.md

The Best Data to Point AI At Is the Data You're Not Allowed to Upload

The files where an AI agent would save you the most time are the exact ones you legally can't paste into a chatbot. That's a wall, not a hurdle.

2026-07-01LOCAL-LLMPRIVACYAI

~/skunk $ cat the-good-enough-solution.md

About 93% of Frontier Quality, for Roughly $0, on a Used Gaming Rig

Two posts named the wall: the files you most want AI help with are the ones you can't upload. So I built the offline answer and graded it head-to-head.

2026-07-01LOCAL-LLMRAGAI

~/skunk $ cat the-cost-wall.md

The Cost Wall: How a Document Agent Torches Your Cloud Budget

Cloud AI feels free because one chat costs a fraction of a cent. Point an agent at a real document corpus and the meter never stops running.

2026-07-01LOCAL-LLMCOSTAI

~/skunk $ cat local-vs-frontier-quality.md

13 Local Models, One RTX 3080, and a 9B That Hits 95% of Frontier

I ran 13 local LLMs on a single 10GB RTX 3080 and scored each one against frontier gold answers. What I found about paying more surprised me.

2026-07-01LOCAL-LLMBENCHMARKAI

~/skunk $ cat imagen-vs-local-images.md

Paying 3× For Image Gen Buys Gloss, Not Correct Text

Part 3 of the benchmark: one hard recreate-this-photo task across all three Imagen 4 price tiers, to see what an extra 3× per image actually buys you.

2026-07-01IMAGENIMAGE-GENERATIONBENCHMARK

~/skunk $ cat health-change-point-detection.md

A Free 6.5GB Model Reproduced a Frontier Blind Read of My Health Data

I gave thirteen local models four months of my own health data with no labels and asked where the trends shifted. One 6.5GB model matched the frontier, in three seconds, for nothing.

2026-07-01LOCAL-LLMHEALTHBENCHMARK

~/skunk $ cat gpu-vs-ram-offload.md

GPU vs RAM: KV-Offload Is a Speed Tax, Not a Bigger Brain

My RTX 3080 has 10GB of VRAM, so I tried spilling the KV cache into system RAM to fake a bigger card. I measured exactly what that trade costs.

2026-07-01LOCAL-LLMGPUPERFORMANCE

~/skunk $ cat hello-from-the-trenches.md

Hello From the Trenches

The real story of what it looks like when you spend your days making APIs talk to each other, and what I've learned in the process.

2026-06-23META

~/skunk $ tail -f updates.log

Stay in the loop

Occasional dispatches from the trenches. No spam.