Builder · TypeScript · Python · Rust · AI tooling · MCP

Master Builder. Making tools that think.

15 years shipping to production. The last few got hijacked by agent-native infrastructure, and honestly I let it happen.

Cofounder at Avistar AI, where we build Astrolayb… an NHI scanner that tells MSPs which forgotten service accounts, IAM roles, and access keys are quietly going to ruin someone's quarter. I also run Petros Industries, an AI tooling shop for the SMBs that can't afford to hire the operations layer they need.

shawnpetros.com · petrosindustries.com · LinkedIn

• Agent-driven build harnesses. Linear-polling daemons that dispatch Claude / Codex agents per issue, with adversarial pre-PR review. Rust where latency matters, TypeScript where the LLM does the work.
• Spec-driven workflow layers. PRDs that translate directly into harness queues. The agent runs the build; the human shapes the spec.
• MCP servers + CLI tools designed for LLM consumption. Memory, retrieval, observability, integration glue. Deterministic, audit-friendly, boring on purpose.
• Production AI for SMBs. Voice bots, content pipelines, workflow automation. The operations layer the buyer can't afford to hire for.

Harness family. Agent-driven build tooling, each piece scoped to a different posture:

• Smithy + Anvil - Reimagined OpenAI Symphony as a Rust daemon. Polls Linear, dispatches one worker per issue, supports both Claude Code AND Codex agents and swaps between them by issue state. Anvil bolts on a new Adversarial Review Linear state between In Progress and Human Review for cross-model audit before any PR lands in front of a human.
• albus-tg-bot - A Telegram surface for a fully-agentic Claude Code session. Voice in / voice out via ElevenLabs, two-way file attachments, CommonMark → Telegram HTML rendering, lock/unlock tool modes, heartbeat + watchdog self-healing pair. Modular TypeScript on Bun, 60+ tests. Couch-coding for the laptop you can't reach.
• Salazar - Autonomous coding orchestrator on the Claude Agent SDK. Planner / generator / evaluator loop with hard validator gates (tsc, eslint, build, test).

Tooling + infra.

• Lithium - Cross-provider LLM-spend aggregator (Anthropic / OpenAI / OpenRouter). Local Rust daemon, SQLite-backed, multiple UI surfaces (CLI, status line, menubar, harness hooks). Mood stabilizer for your AI bill.
• QMD Recall - Deterministic memory recall plugin for OpenClaw. Replaces brittle agent-driven RAG with cited, indexed retrieval. No 20-second pre-answer séance.

Shipped + building.

• claude-code-clone - Minimal Claude Code clone built from scratch. 42 features, zero hand-written code, three agent sessions. The agentic loop is just send → detect → execute → feed back → repeat.
• WindWatts - Open-source wind energy assessment for the National Lab of the Rockies. React + TypeScript + FastAPI. 25 PRs merged.
• PennyBot - Voice bot that joins your Zoom call and speaks. Recall.ai + OpenAI Realtime, built in an evening. MIT.
• ShinyOps - Shiny-hunting planner for Pokémon Legends: Z-A + Scarlet/Violet. Search-first target picker, ranked methods, deep planners for ZA donuts and SV sandwiches. (code)

deAIify ships as both an OpenClaw plugin and a Claude Code Stop hook. The README is a literal infomercial, AS SEEN ON TV. It alone is worth a gander, quite literally unhinged. Also, it really works. Click at your own peril.

Spec-driven development. Tight feedback loops. Real eval suites instead of vibes. CLI tools designed for LLM consumption, not human ergonomics. The unglamorous plumbing that makes agents trustworthy at scale.

📍 Los Angeles · cofounder @ Avistar AI · running Petros Industries · currently building POCaaS and Avistar's agent-native workflow layer in the open.