Applied AI engineer who single-handedly shipped 8 production systems inside a legacy enterprise with no engineering team, no CI/CD, and no code review — then developed a systematic methodology for making that kind of agent-driven engineering reliable and auditable.
The systems below were built by one person with no infrastructure.
Developed and refined across 310+ agentic sessions, 24,000+ conversation turns, and 4 agent platforms (Claude Code, Amp, Cursor, Gemini). Tool-agnostic practices across every project.
AGENTS.md13 governance documents across 7 projects, evolving from 78-line templates to 465-line specifications with MCP security directives, trust ordering hierarchies, and per-server permission tiers. Agent guidance documents are first-class project artifacts.
53% of agentic threads follow a spec-before-implementation pattern. Multi-round QA cycles with reusable audit prompts serve as quality gates — one project went through 5 sequential QA rounds (75% → 100%).
7 distinct safety layers: dry-run mode, JSONL audit logging with rotation, two-phase confirmation (single-use tokens + TTL), filesystem kill switch, 6-tier trust ordering, per-MCP-server permission tiers, and supervisor gates. Trust ordering classifies MCP tool output as untrusted data — a prompt injection defense model.
Persistent inter-session state via MEMORY.md and DEVLOG.md with read-first/write-last protocols. Phase-specific session initialization templates. Multi-phase operational workflows designed as structured campaigns.
33% review/audit · 16% spec/design · 11% implementation
The agent writes the code; I govern the process.
Capital Vacations, 2023–2026. All self-initiated, self-deployed. No engineering team. No dev infrastructure. All currently in production.
Scanner → OCR classifier → SOP-compliant filing. Cleared a 3-month backlog in 3 days.
Power BI + Sheets → SQLite datamart → HTTPS dashboard. 20+ daily users.
Discovered critical dbo-level access across production SQL Server. Documented blast radius, disclosed to CTO/CXO/IT. Converted adversarial situation into remediation partnership.
Gemini Flash transcription → Claude Opus structured notes. In production.
8 min/tour → 2 min. Deployed to regional staff.
Python/pyodbc with mock mode, audit logging, pre-flight checks, explicit transactions.
Google Sheets replacing legacy tooling. Daily driver across departments.
Automated guest profile compilation across 4 systems.
Architecture patterns: two-phase confirmation, JAIL_ROOT security boundaries, audit logging, dry-run modes, AGENTS.md governance.