Marketer who builds.

Red Line Works: a full brand launch in one day

A firefighter-owned moving and dumpster company in Augusta with a real reputation and no way to be found: no website, no domain, an unmanaged Google listing, and an early-AI logo. Customers who heard about them couldn’t find them to hire them.

In one working day I shipped the fix end to end: a rebranded vintage badge logo, a custom site live at redlineworksllc.com with HTTPS, click-to-text, and local-SEO schema, a Google Business Profile overhaul, branded email on the domain, and a review engine: direct review link, QR codes for the dumpsters, and print-ready card designs.

Claude wrote the code and produced artwork. I owned everything that made it work: the strategy (reviews and Google first, dumpsters featured), three rounds of brand direction with the owners, account recovery, DNS and deployment, and every quality call in between. Paying client, delivered same day.

The ICLR paper that would not fit on a GPU

An ICLR 2026 paper ships a method for stripping AI-cliche language out of language models, and its released fine-tuning code ran out of memory on a 141GB card, one of the largest single GPUs money can rent. The obvious move is bigger hardware. The overnight run was scripted as a fallback chain instead, and by morning the data said something better: halving the model adapters moved peak memory by half a percent, so the adapters were not the problem.

That reasoning led to a two-line fix that cut peak memory from 139.6 GiB to 28.3 GiB, about 5x, letting the same configuration converge on a card with 57% of the memory at a third of the hourly cost. I also verified the paper's core claim independently: across 2.5 million generated words on two model sizes, the method cut AI-cliche language 73.7% and 68.5%, scored against a yardstick chosen separately from the pipeline's own ban lists. The fix is published in a public repo and filed upstream, and the paper’s lead author confirmed the finds the same day and asked for a pull request. That PR is open now.

Claude wrote the code and ran the cloud GPUs. I owned the experiment design, the diagnosis logic, and the call that mattered most: prove the fix live before publishing a word of it.

The pool inspection app

Inspections used to be handwritten: printed sheets with one-word labels and a blank line under each. No photos, no structure, nothing you’d be proud to hand a customer.

So I built a web app techs open on their phone at the pool. Photos shot inside the form. Structured condition checks, equipment pass and fail, notes. One tap generates a branded PDF report, saved to the device and emailed to the customer on the spot.

Built solo, with Claude writing code while I owned every product decision: the fields techs actually need, the order they work in at a real pool, and what a customer wants to see. Deployed on Netlify and adopted by the field team.

Instagram, from zero

On February 16, 2026 the company Instagram went live with a fully populated grid. Ten-plus posts, zero placeholders. Strategy, photography, design, and copy produced end to end, then run as an ongoing channel alongside Facebook.

The 50-company teardown

I analyzed the content of 50 pool service companies to find what separates the winners. The answer was consistent: authentic field photography beats template graphics.

That finding became the company’s content direction, a 45-page content marketing strategy, and a 17-page operations playbook that turned strategy into platform-specific formats, posting cadence, and editorial themes.

The second brain: two years of research, distilled by agents

Like most marketers, I saved everything and reread nothing: 591 posts collected over two years, a research graveyard. So I built the machine that reads it for me.

A harvester pulled every saved post plus its top-voted comments (4,301 of them, because the comments are where posts get fact-checked). Then a multi-agent pipeline classified all 591, distilled the keepers into a knowledge base where every entry cites its sources and carries a confidence level, and merged the duplicates. The junk filter auto-rejected 75 disguised ads, including astroturfed threads where vendor accounts praised their own products.

AI ran the agents. I owned what made it work: the pipeline design, the topic taxonomy, the quality bar, and the skepticism rules. Now anything new I read gets checked against the index first: it tells me what I already know and files only what’s new.

Don’s Dock Seafood Market

A 24x36 species reference poster of 34 Gulf fish for a seafood market in Madeira Beach, Florida. Brand identity, grid system, and AI-assisted image production end to end. Now in print production.

The Marketing Workbench: six tools, one discipline

Every small marketing shop bleeds hours on the same jobs: the monthly client report, the A/B test call, the review replies, the ad-fatigue check, the landing-page critique, the content repurposing. The AI tools sold for this have a worse problem than the hours. They invent drama, and a marketer who ships an invented number once loses the client.

So I built six tools to one discipline. Each is a single HTML file that works from a double-click, no install, no server, your own API key stored only in your browser. Parsing, math, and verdicts happen in plain, readable code before any AI writes a word. The A/B judge's verdict is deterministic statistics, never a model's opinion, and every tool shows you the verified numbers before a single sentence is generated. Each ships with planted trap data whose correct answers were written down before the tool was built, so the tool is graded against the data and never the other way around. 543 harness checks across five test harnesses, all green, plus a byte-parity proof that the report studio's brain matches the tested pipeline it came from.

Claude wrote the code. I owned the discipline: which jobs deserved a tool, the honesty rules the writing runs under, the trap data, and the bar that nothing ships to a real marketer until I have run it live myself.