I build and operate multi-agent AI systems, keep technology working for the people who depend on it, and I'm documenting a deliberate move into cybersecurity — in public, including the parts that don't go smoothly, because troubleshooting is the job.
Self-directed engineering projects — designed, built, tested, and independently verified.
Problem. A single AI agent asked to deliver a whole project skips steps, grades its own homework, and fails silently. Production-quality delivery needs planning, approval gates, execution, and verification by separate parties — the same separation of duties a real engineering organization uses.
Architecture. A three-level agent hierarchy — one Grand Strategist (orchestrator), three domain Advisors (architecture / implementation / verification), and tiered execution sub-agents routed by cost-of-being-wrong across frontier and efficiency-class LLMs. Nine governing laws: chain of command, plan-before-action with explicit approval gates, independent verification, halt-don't-guess escalation, two-strike tier escalation, step budgets, and a files-are-truth audit trail. Built on Model Context Protocol (MCP) with structured JSON-schema tool calling; every mutation is git-checkpointed. Packaged as a reusable skill with exportable prompt packs for Claude Desktop and the Hermes agent ecosystem.
Implementation. Deployed as a production skill and validated on an end-to-end calibration build: 11 gated tasks, every one through plan → advisor approval → gated execution → independent audit, with a complete on-disk paper trail (plans, approvals, verdicts, results, flight log).
Key results (verified).
Problem. A local multi-agent ecosystem (Hermes) — formerly Hermy Command Center — had no single pane of glass — agent states, execution traces, task queues, and runtime telemetry lived in scattered terminals, so a stalled agent could burn compute unnoticed.
Build. A centralized mission-control interface — agent roster and hierarchy view, live activity feed, approval inbox for gated actions, task tracking, and inference telemetry — engineered as one self-contained HTML/CSS/JavaScript application (~100 KB, zero dependencies). Supervises local model runtimes (Ollama, LM Studio), balancing latency, privacy, and compute cost. Dark-first UI with a custom pixel design system.
▶ Open the working interface — this is the actual dashboard, not a mockup.
To prove the orchestration system on real work, The Tree built a password-strength analyzer end-to-end: three sealed spec documents, a scoring library (five criteria: length, character variety, uniqueness, common-password matching, repetition/sequence penalties), a CLI, and a 248-test suite.
Why does "password" score 0/100? It takes the maximum dictionary penalty, has one character class, and gains too few length points to recover — exactly what the published scoring spec predicts. Clean exit codes (0 success · 2 usage error), deterministic output proven across processes.
The audit gate failed the specs on the first pass — two real defects, fixed and re-verified before implementation began. That failure is the system working.
Current build: a synchronized knowledge fabric (Obsidian Sync vault) shared by Claude Desktop, a Hermes cloud agent (Nous Portal), a local Hermes agent, and the Command Center — so every agent reads the same project files, decisions, and memories. Model routing by cost-of-being-wrong: a frontier orchestrator, an independent advisor model, and efficiency-class sub-agents. Documented as it's built.
A deliberate, documented pivot — labs in public, credentials in sequence, employer-sponsored.
| A+ Core 1 (220-1201) | ● IN PROGRESS |
| A+ Core 2 (220-1202) | ○ QUEUED |
| Network+ (N10-009) | ○ QUEUED |
| Security+ (SY0-701) | ○ QUEUED |
Sequenced deliberately: Security+ assumes networking fluency it doesn't teach — I'd rather hold the credential and the competence at the same time.
Organized the way recruiters and applicant-tracking systems actually read.
Eighteen years of making complex things work — international education, records research, K-12 technology support, facilities operations, and consumer technology.
Role concluded when Washburn Elementary closed under district budget constraints.
Formal coursework, professional certifications, and current programs.
Employer-sponsored (Target). 33+ hands-on labs across firewalls, VPN, directory services, encryption, vulnerability assessment, and log analysis.
Sequenced so the networking foundation is in place before the security credential.
Regulated-facility operations: water-chemistry testing, compliance logging, and public-health code adherence.
Teaching English to Speakers of Other Languages; foundation for a decade of instructional design and end-user training.
Coursework: Networks & Server Structures · Human-Computer Interaction & Usability · Web Standards Design · Designing & Advanced Computer Graphics · Intellectual Property & Law · Principles of Production Process.
Microsoft Word & Microsoft Excel certification coursework (ADMS 1260 / ADMS 1265).
A decade of self-taught game development, animation, design, and audiovisual production — the systems-thinking foundation under the AI work.
Original social-media brand graphics for YouTube, Twitch, and Instagram — Adobe Photoshop and Illustrator: hand-drawn assets, texture work, and cross-platform layout.
Open to technical operations, IT support / help desk, AI systems support, and security-track roles — Minnesota (Twin Cities) or remote.