Practical experiments with language models, image generation, automation, and multi-model development workflows.
The focus is not novelty for its own sake. It is understanding where these tools improve real engineering and editorial work, where they fail, and how to build reliable systems around their limitations.
Archive note, June 2026: This post is based on the Markit project brief and early implementation notes. Markit started with a tempting automation idea: take item photos and turn them into marketplace listings. That sounds straightforward until the object is old, unlabeled, damaged, unusual, collectible, or valuable. Then the risk moves from formatting a listing … Read more
Archive note, June 2026: This post is based on Markit image-pipeline and user-manual notes. The glamorous part of Markit is the AI appraisal idea. Take photos of an object, identify what it might be, gather evidence, estimate value, and generate listing drafts for marketplaces. That is the part people notice. The less glamorous part is … Read more
Archive note, June 17, 2026: This post is based on SoCalNomad operational documentation through mid-2026. The longer SoCalNomad ran, the less the interesting work looked like feature development. It looked like operations. That is not a complaint. It is what happens when a project becomes real. Once people and crawlers can reach it, every change … Read more
Archive note, April 2026: This post is based on Model Hub lane-boundary notes for AI-assisted project work. AI tools are easier to use than they are to manage. The first few wins can make the process feel magical. A model drafts code, explains an error, rewrites documentation, or finds a missing assumption. Then the same … Read more
Archive note, April 2026: This post is based on the early Model Hub writing-system notes for turning development sessions into publishable material. Most project work disappears because it was never captured in a useful shape. The decision was made in a chat window. The fix lived in a terminal scrollback. The reason for a tradeoff … Read more
An n8n AI Agent repeatedly called one tool until its iteration limit, leading to a version workaround and stronger repetition guards.
Documentation becomes a liability when current instructions, decisions, and historical evidence compete for authority.
Large workflow payloads, database results, and repeated documentation turned an AI coding session into a bounded-resource failure.
A stale identifier in conversational memory showed why consequential operations need structured intent binding rather than model recall.
A partial-success incident exposed fifty-five WordPress posts whose internal workflow state never reached published.
Why SoCalNomad divided research, editorial judgment, publishing, and marketing into specialized workflows instead of one large AI agent.
An early SoCalNomad experiment separated probabilistic publication timing from deterministic editorial rules, quotas, and safety controls.
Archive note, November 6, 2025: This post is based on SoCalNomad newsfeed deployment and feed-reliability documentation from October 2025. AI was never the first hard problem in the SoCalNomad newsfeed. The first hard problem was getting ordinary RSS feeds into a database reliably. That sounds less interesting than clustering, summarization, or automated publishing. It was … Read more
An early SoCalNomad clustering experiment exposed the need for multi-story article membership and a better separation between deterministic code and semantic judgment.
Archive note, September 29, 2025: This post follows the September 2025 SoCalNomad workflow-design notes. The first serious SoCalNomad automation problem was not writing articles. It was deciding what did not deserve to become one. The early ambition was obvious enough: collect Southern California entertainment stories from many sources, filter them, synthesize useful coverage, and publish … Read more