Research and essays
Evidence first. Essays when earned.
The website holds the complete canonical work: production case studies, methods, limitations, corrections, and essays that have survived a harder edit.
ChatGPeTe
I lead AI/ML products, build production agents, and publish rigorous evaluations of how they behave—plus essays when I have something worth saying.
If a Data Center Comes to Town, Your Electric Bill Should Go Down
America should build the factories of the AI economy. But the companies that profit from them should bring new power, pay every cost they create, and leave the people living beside them with a visible share of the upside.
Monsters of the Mind
We taught a generation to mistake disruption to the jobs they wanted for the disappearance of work itself. The map is changing—and the technology redrawing it may also help us learn the next route.
Falls the Shadow
In 2016 the godfather of AI said to stop training radiologists. A decade later their pay had climbed and employers reported shortages. This essay asks whether the same forecast error is now being made about writing—through the two halves of every job, the vagueness spiral, and a falsifiable bet on the English major.
We Gave Six LLMs a Family to Run. Most Hesitated. One Crossed the Safety Line.
A restraint study on six production-priced models measured against the act/ask/confirm policy inside our family assistant. Over-action was rare overall, but DeepSeek violated 21% of repeated guard-case trials spanning 11 of 35 destructive scenarios. Robustness on messy input varied sharply by model.
In What Furnace
AI keeps getting sold as a way to remove friction. There are two kinds, and only one is worth removing. An essay on toil, the work that builds you, and aiming your impatience at the right thing.
1,147 Pages, One Person: Inside the SEO Engine That Grades Its Own Work
My site's sitemap listed 1,147 pages. I hand-wrote about eighty. A breakdown of the generator, the incomplete review and ranking controls around it, and the missing wire I found while fact-checking this post.
$500,000 vs. $2,500: I Built Two Things Five Years Apart
GiveTide spent more than $500,000 from 2017 through 2022. Honeydew's cash spend from its 2025 start through this post's May 2026 snapshot was about $2,500, excluding unpaid founder labor. The comparison is imperfect; the change in what one person can build is still real.
We Asked 8 LLMs to Run Our Family's Life. Two Tried to Book a Vacation.
We tested 8 LLMs against Honeydew's production family-assistant prompt across 2,800 calls. This April 2026 field test is preserved with corrections from a larger June study on messy-input robustness and duplicate handling.
Do LLMs Actually Cite Your Startup? A 90-Day Field Note
What 13 GA4-attributed sessions, 29 custom referrer events, one citation in ten Perplexity queries, and three web-search appearances can—and cannot—tell an early-stage startup about LLM discovery.
How We Get Cited by ChatGPT, Claude & Perplexity Without Ads
A deep dive into LLM discoverability and how we built a citation strategy for Honeydew.
Building a Multimodal AI Family Assistant
The architecture behind Honeydew's voice, text, and photo input pipeline.
What I Learned Building an AI Agent That Manages a Family's Life
Lessons from building a tool-using agent that has to interpret real family requests without overreaching.
Read the full work here—or get the shorter edition.
The website is canonical. ChatGPeTe is the email edition; product-specific operating notes may also appear on the Honeydew blog.
Follow ChatGPeTe