I run production systems for a living and build small businesses on the side. Most of what I make these days has AI somewhere in the loop — as a co-author, a co-developer, or the thing I'm operating in production. East Texas.
Production experience deploying and operating LLM-based agents at enterprise scale. SRE discipline applied to non-deterministic systems — observability, failure modes, rollback paths, cost discipline.
Five live web properties built with AI as a co-developer. Content pipelines, programmatic SEO, embedded LLM tooling, and the unglamorous plumbing that turns a model into a product.
Notes on operating AI in the real world — what breaks, what doesn't, and where the SRE discipline does and doesn't transfer. (Writing surface coming soon.)
Not a chatbot. Not a productivity app. Chief of Staff is the operational layer that runs between me and everything else — daily briefings, task triage, decision support, and context management across the full indie portfolio. The architecture is deliberate: persistent memory backed by a secured Supabase connection means context actually survives across sessions. No third-party sync, no ambient logging, no model training on my data. Built as a PWA so it installs and behaves like a native app. The whole premise is that if you're going to trust an AI with how your day runs, the plumbing underneath it should be something you designed and own.
SITREP is a personal OSINT terminal — a daily situational-awareness tool that aggregates signals I actually care about and surfaces them in one place. Threat feeds, infrastructure noise, news with relevance filters. The SRE instinct applied to information: if you wouldn't run a production system without dashboards and alerts, why would you run your day without them? Built as a PWA, deployed privately, not a product.
Two sister sites built around a shared editorial philosophy: take ideas seriously, write for curious people who aren't academics, and never oversell the mystery. Frequency Unknown covers consciousness, perception, psychedelics, and the philosophy of mind — the honest, readable translation of ideas that usually live behind jargon or breathless woo. The Gnostic Guide covers early Christian mysticism, Nag Hammadi texts, and gnostic philosophy with the same rigor: sourced, skeptical, and written to respect the reader's intelligence. Both are Astro, both on Vercel, both with active content pipelines and newsletter infrastructure. 25+ published articles between them as of mid-2026, with 20-week content plans in motion. Built because the subject matter deserved better treatment than it was getting online.
Personal branding tools assume you already know what your brand is. Most people don't. yourEra starts a step earlier: a voice interview where Claude listens to how you actually talk — your phrases, your humor, your rhythm, your weird hangups — and turns ten minutes of speech into a brand brief and thirty days of content that sounds like you wrote it. Not like ChatGPT wrote it. The whole thing is a multi-agent pipeline under the hood, but what users feel is simpler: they spoke for ten minutes and got a month of posts back. $19/month. Built in a week. Live in production.
An editorial site for people on GLP-1 medications — the demographic that's exploded in the last two years, written for and largely ignored by mainstream health publishing. WordPress on the surface, AI-assisted content pipeline underneath: Claude in the loop for research, drafting, citation verification against actual sources, image production, and SEO. The Pinterest visual system is the part I'm proudest of — a typography-first identity that reads as a real publication, not a content farm. Built one article at a time with the discipline of a magazine, not a blog.
HTS classification, risk analysis, and action plans in 30 seconds for small importers who can't justify a five-thousand-a-month enterprise tool or a 300-an-hour customs broker. LLM-backed, Stripe-monetized, paired with a Beehiiv newsletter for habit-forming touchpoints. The product is — honestly — what a good first answer from a good AI looks like, packaged so a non-technical operator can actually use it.
A state-by-state matrix of benefits eligibility (SNAP, Medicaid, unemployment, and more), generated and maintained through a templated pipeline rather than hand-written page by page. One template, scripted AI assistance, dozens of pages — the kind of programmatic content surface that makes sense when the underlying data is structured but the audience needs prose.
Two lightweight, legal-adjacent reference tools. Statute-of-limitations lookup by state and claim type; child-support calculators by state with formula transparency. AdSense-monetized, fast, and built to look like real publications rather than scraper farms — same editorial discipline as the bigger sites, smaller surface area.
AI is a co-author and a co-developer, not autopilot. The things I ship are reviewed, sourced, and fact-checked before they go out — outbound citations get verified against the actual source, not pulled from memory, because the model is confident even when it's wrong and the only fix is a human checking.
The system around the model matters more than the model. I keep persistent context files, named lessons, structured session protocols, and a daily log that survives across conversations. Most of what looks like “the AI did this” is really the scaffolding I built around it.
SRE discipline transfers. Observability, blast-radius thinking, rollback paths, cost discipline, postmortems — the same operating habits that keep production systems from cratering also keep AI systems from quietly drifting into nonsense.
Best reached by email: hello@byron-walker.com.