Byron Walker
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.
Flagship build
An incident-response platform that detects, remediates, and grades its own work
A real containerized fleet, real injected faults, and real remediation — actual
docker restart on actual containers,
not a simulation of one. The system triages with an LLM, fixes the fault, then grades
the fix against ground-truth metrics rather than model opinion, and proposes tuning
to its own detector when it notices it's noisy.
Postgres state machine enforced at the DB level · LISTEN/NOTIFY event bus · pgvector case memory · streaming replication, primary → standby · self-tuning gated behind human approval. Runs as a local lab on Docker — a portfolio artifact, not a hosted product.
Type incidents in the terminal above to replay a live incident lifecycle.
What I work on
AI agent operations
Production experience deploying and operating LLM-based agents at enterprise scale — and at home, an autonomous incident-response platform that runs itself. SRE discipline applied to non-deterministic systems: observability, failure modes, rollback paths, cost discipline.
Indie AI builds
Seven 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.
Systems thinking, in writing
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.)
Now
Selected work
incident-platform Autonomous, self-improving incident response
The thesis: take the incident-response work I do at enterprise scale and rebuild it as an autonomous system that improves itself. A containerized fleet reports real metrics; an injector applies real faults (CPU pegging via stress-ng, paused containers); a detector opens incidents; workers triage with one Haiku call, match runbooks, and execute real docker restart / unpause commands. Then the part most AI-ops demos skip: a postmortem agent grades every resolution against ground-truth metrics — not LLM opinion — writes vector-embedded case memory, and proposes bounded, auditable tuning to its own detector. The proposals are human-approved by design; the system does the hard part of noticing its own failure modes.
The debugging stories are the real portfolio: a detector that was structurally blind to hosts going dark (a metric-breach detector can't see absence of signal — fixed with liveness heartbeats), dashboard metrics that cried wolf on an idle system, and a timestamp that lied for three days because Postgres now() is transaction-scoped — every incident was backdated by the quiet period before it. Finding that took systematically falsifying every other hypothesis first.
chief-of-staff Claude-powered personal AI operating layer
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. Recently put through a full security hardening pass — signature-verified auth on every endpoint, user-scoped queries and mutations, encrypted server-only Google tokens, and human-in-the-loop email triage so untrusted content never writes to the database without a tap. 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 OSINT situational-awareness terminal
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.
thefrequencyunknown.com · thegnosticguide.com Independent editorial network
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.
yourera.app Voice-powered personal brand builder
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.
supplementsforglp1.com GLP-1 supplement editorial site
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.
benefitsmatch.net Programmatic SEO for benefits eligibility
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.
statutelookup.net · supportcalc.net Reference utilities
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.
How I work with AI
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. The incident platform above is that idea taken literally: the system grades itself against ground truth, because trusting a model to grade its own homework is how ops automation drifts into fiction.
Contact
Best reached by email: hello@byron-walker.com.