I build automated systems that run themselves — and businesses that run on them.
27 years shipping production software, the last several building AI-native systems that run real businesses without me in the loop: content, real estate, e-commerce, and publishing. Mostly solo, in Python and Go, live on a Linux VPS, unattended by design.
Everything below is in production right now.
17 years continuous operation · 110% sales lift in 6 months · 150–300 qualified leads/week · Open-source tools other builders adopted on their own
Selected Work
Governator — governance layer for coding agents
A Go supervisor sitting above coding-agent CLIs (Claude Code, Codex, OpenCode, Pi) that turns file scope, command authority, and budgets into deterministic checks around every agent run. No tool call executes ungoverned: a separate validator and merge gate decide whether work ships, not the model, and OS-level containment (Landlock, network namespaces, cgroups) backs those guarantees at the kernel instead of trusting the backend’s word.
It is also a cost-control layer. Output compression and YAGNI-biased prompting keep diffs small, circuit breakers route away from a failing provider instead of burning tokens on doomed retries, quota-aware routing avoids backends about to hit a cap, and a hard spend cap halts a run before it becomes a surprise bill. GitHub
Assayer — verification layer for AI output
Verify, quarantine, trace. Every generated result is checked against configurable rules (schema, length, boilerplate, placeholder text, language sanity, dedup) before it is allowed downstream; anything that fails is quarantined for review instead of shipped. A durable local outbox means the evidence survives even when the database write does not.
It is a plain library call, so it works the same on an interactive agent’s output or an unattended API pipeline, and its pass/fail history feeds back into Governator’s routing — a backend’s real output quality, not just its uptime, decides what gets picked next. It runs on every pipeline I operate that touches an LLM. GitHub
Zone31 — AI divination platform
A multi-tenant consumer platform where an oracle answers in the voice of one of thirty-seven gods, sibyls, and adversaries, with tarot, dream interpretation, I Ching, astrology, ritual, and shadow work beside it. Every voice is compiled from a source-controlled “voice bible” rather than hand-edited prompts, so character stays consistent across surfaces and releases.
Under it sits the infrastructure that makes a public LLM product survivable: a safety and crisis layer, per-day cost governance across every model call, a provider chain that fails over between vendors mid-request, and point-in-time database backups. Built for the audience Tarotsmith has been serving since 2009. zone31.com
Production Pipelines
Tarotsmith Network — 17 years of continuous, revenue-generating operation (2009–present), and one of the longer-running AI-assisted content operations of its kind. Multi-agent article generation, weekly horoscopes, automated publishing. tarotsmith.com
Sawasdee Realty — An automated lead-gen system producing 150–300 qualified Bangkok leads a week: Python scraping, AI scoring, and multichannel distribution across TikTok, property portals, and Line OA. The work most agencies still do by hand. sawasdeerealty.com
Cookie Pipeline — One photo in, five platforms out, no humans in between: photo drop → AI script → Thai TTS voiceover → vertical video → WordPress, TikTok, Instagram, Facebook, and YouTube.
MonPaga / Waking Cup — Review aggregation and e-commerce automation that drove a 110% Lazada sales increase in six months for a Thai marketplace seller with no in-house tech team.
Personal AI Tools
Built for myself first, published because other builders asked.
second-brain-interviewer — A Telegram bot that interviews you by voice. A psychoanalytic AI character works through nine categories (values, fears, mental models, working style, history) and writes your answers straight into a personal second brain. State persists across sessions; roughly $0.50 for the full interview.
second-brain-brief — A daily brief that queries a Supabase vector store, synthesizes with Gemini Flash, and lands in your inbox before you open it. The second brain reports to you, not the other way around.
compass — Four questions, once a day, and AI synthesis returns a single heading for the session. A calibration ritual, not productivity theater.
future-self-letter — Claude writes a letter from your future self, sealed for 90 days and delivered by Telegram on schedule. It forces the question: did you do what you said you would?
Personal Podcast Network — Eleven private shows on one engine, covering every day of the week: Gemini Flash script generation, multi-speaker TTS, Cloudflare R2 hosting, private RSS. No browser automation, no third-party services, fully unattended. The shows range from a digest of your own YouTube and podcast subscriptions, to esoteric history, depth psychology, UAP disclosure, AI/builder news, and curricula distilled from hundreds of book reviews — but the interesting part is the engine underneath, not the topics. personal-podcast-engine is the shared base class: subclass it, implement fetch_sources(), get a full episode from any input.
Toolchain
Orchestration: Claude Code · Codex · OpenCode · Pi
Model APIs: OpenAI · Anthropic · Qwen · DeepSeek · GLM · Minimax · Groq · OpenRouter · Fal
Agent governance: Five tools I run on every agent session, because the combination compounds — gate, verify, compress, minimize, repeat. Governator gates every tool call before it runs and Assayer verifies the output before it ships; RTK compresses command output 60–90% on the way to the model, Context Mode keeps raw tool output out of the conversation entirely by processing it in a sandbox, and Ponytail injects a YAGNI-first ruleset so the agent reaches for existing code and the smallest diff before writing anything new. That is the difference between a multi-agent pipeline that is affordable to leave running and one that quietly burns the budget. OmniRoute sits alongside them, routing one endpoint across 290+ providers with quota-aware fallback so a rate limit never stalls a run.
About · Experience · Contact · GitHub