I build things with AI coding systems — Claude Code, Codex, and the rest — and I'm good at getting real work out of them.
Security research, production websites, chatbots, automation, bug bounty hunting. Different domains, same skill: knowing how to drive these tools, where they break, and what to do about it when they do.
📍 Tillamook, OR · Sacramento, CA · open to remote · mtarcure888@gmail.com
- A critical-severity vulnerability — found, triaged and paid through a HackenProof bounty program.
- A production website for a paying client, plus builds for local businesses.
- Chatbots and automation wired into the tools people actually use.
- Open-source plugins for Claude Code, used by other developers.
claude-vibe-squad · Python · AGPL-3.0
Everything above kept hitting the same walls: agents losing the thread mid-task, work that looked finished but wasn't, no way to check one model's output against another. So I built the tooling to fix it, and I use it every day — for client work, personal projects, and bounty hunting.
Multi-model orchestration. One coordinator routes scoped Markdown task packets across 4 frontier model families — Codex, Claude, Gemini and Kimi — to 68 role-based specialists, with capability-aware task routing that binds each specialist to the model that suits it.
Sandboxed execution. Every task runs in an isolated git worktree with an explicit write scope. A worker can't touch what it wasn't granted, and nothing merges without passing its gate.
Cross-family evaluation. Work is reviewed by a different model family than wrote it — LLM-as-judge with an anti-affinity rule, so a model never signs off on its own reasoning.
Durable memory with a closed learning loop. Findings are recorded as candidate notes, recalled into later work, and promoted to verified only when a task that used them passes review. Memory that isn't used doesn't get trusted.
Context engineering and drift prevention. Bounded resume capsules instead of dumping state, thread charters that hold the original ask across topic switches, and skills that fire on trigger rather than on request.
Built from things that actually went wrong. A validator that reported "0 errors" for weeks while a bug meant it checked nothing. A memory loop that silently stopped learning for a month because unit tests couldn't see a runtime defect. Config that guessed instead of refusing. Every gate in it exists because something got past the last one.
| claude-code-council | 5 AI advisors across 3 model families with anonymous peer review — for decisions where being wrong is expensive |
| gemini-review-plugin | Adversarial code review on a second model family, structured JSON output, no CLI dependency |
| chain-strike | Turns separate medium-severity findings into one demonstrated critical — exploit-chain construction for authorized bug bounty work |
Claude Code · Codex · Python · TypeScript · MCP (Model Context Protocol) ·
multi-agent orchestration · agentic workflows · task routing · durable memory ·
context engineering · LLM-as-judge evaluation · guardrails & human-in-the-loop gates ·
security research · AWS Certified AI Practitioner · Google AI Professional Certificate
Open to remote work — AI operations, agent engineering, evaluation, or anywhere getting real output from these systems is the job.