"I don't build systems to impress β I build them to quietly endure.
From bare-metal infrastructure to AI agent governance: solid systems begin with steady hands."
With 11+ years in IT Operations & Systems Engineering (bare-metal virtualization, multi-site enterprise networking, and zero-downtime disaster recovery), I approach Artificial Intelligence from a First-Principles engineering perspective.
While most treat AI as prompt-driven magic, I treat it as an untrusted, high-velocity compute layer that demands:
- Deterministic Guardrails: Hard runtime boundaries, process sandboxing, and write receipts instead of hoping the LLM "behaves."
-
Epistemic Integrity: Enforcing a strict separation between hypothesis (
[HYPO]) and physically proven evidence ([PROVEN]). -
Zero-Root Infrastructure: Multi-layer perimeter defense (Edge WAF
$\rightarrow$ Traefik v3$\rightarrow$ CrowdSec$\rightarrow$ Passkey FIDO2) with bounded autonomous self-healing.
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β COGNITIVE LAYER: Governed Second Brain β
β β’ SSoT Knowledge Base & Deterministic Orchestrator (wiki.py) β
β β’ ADLC Finite State Machine & Gate D Deep-Why Verification β
ββββββββββββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββββββββββ
β Issues Path-Scoped Write Receipts
βΌ
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β EXECUTION LAYER: g8s (The Gatekeepers) β
β β’ Zero-Trust Go Binary Runtime for AI CLI Workers β
β β’ Capability Dropping, Subagent Sandboxing & Supervisor Fix Loop β
ββββββββββββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββββββββββ
β Controlled Action / Code Changes
βΌ
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β OPERATIONS LAYER: AI-Assisted Homelab Platform (IaC) β
β β’ Bare-Metal Proxmox VE, Rootless Podman, Tailscale Zero-WAN Mesh β
β β’ Autonomous SRE Butler Agent: CTI Ingestion, SBOM & LogSQL Hunting β
β β’ Human-in-the-Loop (HITL) HMAC-SHA256 Telegram Approval Cards β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
π‘οΈ g8s (The Gatekeepers)
A Lightweight, Zero-Trust Process Execution & Capability Harness for AI Agent CLI Workers
βk8s orchestrates your compute containers; g8s orchestrates your AI subagents.β
- Two-Tier Agent Harness: Enables high-tier Brain orchestrators (Claude 3.7 / Opus / GPT-4o) to safely dispatch mechanical tasks to lightweight CLI workers (Antigravity CLI, Claude Code CLI, Gemini CLI) under strict sandbox constraints.
- Cryptographic Write Receipts: Grants time-limited, path-scoped write authority (Schema v3) to prevent unauthorized file mutation and codebase drift.
- Supervisor Fix Loop: Automated execution verification, metrics ingestion, and self-correcting agent loops.
Production-Inspired Reference Architecture & SRE Automation for Bare-Metal Homelabs
-
3-Tier Balanced Architecture: Perimeter Ingress & Identity (Traefik v3, CrowdSec, Tailscale Mesh, Dual-Tier IAM)
$\rightarrow$ AI Gateway & Operations Plane (LiteLLM, SRE Butler Agent)$\rightarrow$ Observability & GitOps Core (VictoriaLogs, VictoriaMetrics, Age Encryption). - Dual-Tier Zero-Trust IAM: Biometric Passkey FIDO2 via Pocket ID combined with ForwardAuth proxy via Authelia and LLDAP.
-
Closed-Loop SRE Automation: SRE Butler periodically polls CTI feeds (CISA/GHSA), performs
$O(1)$ local SBOM matching, executes retroactive forensic hunts via VictoriaLogs LogSQL, and dispatches HMAC-SHA256 signed Telegram approval cards (300s TTL). - Zero-Spam GitOps Engine: Automated hypervisor drift reconciliation every 15 minutes with content-addressable SHA-256 suppression and Age Zero-Knowledge secrets encryption.
π‘οΈ defense-in-depth
Governance Middleware Between AI Agents and Your Codebase
- Acts as an automated defense layer preventing AI agents from generating hollow artifacts, executing unverified cowboy commits, and polluting Single Sources of Truth (SSoT).
- Enforces strict contract schemas and validation checks before mutations reach Git repositories.
π€ ContribAI (Contributor β original by @tang-vu)
Autonomous Open Source Contribution Engine
- Contributed to autonomous multi-agent pipelines for discovering open-source issues, synthesizing reproducible tests, and authoring pull requests.
- 9 PRs successfully merged across 21+ open-source repositories.
(Architectural blueprints and private self-hosted fleets operated under strict Zero-Leakage OpSec standards)
- Challenge: Scaling personal knowledge and technical decision-making across dozens of AI agent sessions without context drift, unverified assumptions, or knowledge rot.
- Architecture & Mechanism:
- Engineered a deterministic Python runtime (
wiki.py) functioning as a physics engine for the vault: forces orientation, JIT similarity searching, and epistemic labeling before any file modification. - ADLC (Agentic Development Life Cycle): Enforces finite state machine (FSM) transitions and Red Test Proof sections on engineering plans, backed by automated Git pre-commit gates.
- Memory Distillation Pipeline: Background daemon scanning execution logs to distill ephemeral agent interactions into proven axioms with Zero DLP leakage.
- Engineered a deterministic Python runtime (
- Impact: Zero hallucinated rules across 500+ documentation nodes; 100% auditable knowledge lineage.
- Challenge: Managing a high-availability bare-metal virtualization cluster hosting 20+ services with zero public exposure and sub-second incident triage.
- Architecture & Mechanism:
- Zero-WAN Exposure Law: Management interfaces (Proxmox VE, RouterOS, VictoriaLogs) strictly restricted to internal LAN and Tailscale Mesh VPN (Zero-Trust Encrypted Mesh).
- Dual-Plane Observability: Synthetic uptime monitoring separating compute plane (
[LAN]internal container health) from edge ingress ([WAN]public domain health), isolating routing vs service failures in under 2 seconds. - Single Authority DHCP: Centralized RouterOS static lease binding as the single source of truth for container network identities.
- Impact: 99.9% fleet availability, zero unauthorized ingress attempts, and under 5-minute automated disaster recovery.
π€ AI Agents & Governance
- π‘οΈ g8s (The Gatekeepers) β Zero-Trust process execution & write receipt harness for AI CLI workers
- π‘οΈ defense-in-depth β Governance toolkit for AI coding agents
- π§© mcp-orchestration-framework β Multi-agent SoT bootstrap PoC using Model Context Protocol
- β¨ awesome-prompt-engineering β Curated prompts and guardrail patterns for AI + DevOps automation
- π€ ContribAI (fork/contributor) β Autonomous OSS issue discovery and PR contributor (9 PRs merged)
ποΈ Infrastructure, SRE & IaC
- ποΈ IaC β AI-Assisted Infrastructure Operations Platform & SRE automation for bare-metal homelabs
- π foss-virtualization β Open-source guide to virtualization & containerization
π οΈ Utilities & Tools
- πΆ tuneflow β Elegant, elderly-friendly YouTube audio downloader & player
- π» cmdToolForHelpdesk β Windows IT support automation toolkit (batch scripts)
- π hash-checker β Cross-platform Rust hash checker (CLI + GUI)
- π proxy-checker β Proxy availability checker with Telegram reporting
- π₯ awesome-download-manager β Curated download manager reference guide
A decade of bare-metal and cloud operations providing the bedrock for scalable AI systems.
| Domain | Technologies |
|---|---|
| Systems & Languages | |
| AI Agent & Protocol Systems | |
| Virtualization & Containers | |
| IaC & Automation | |
| Security & Networking | |
| Observability & SRE |
- Mechanism over Prompting β Don't instruct AI to be careful; build guardrails and process harnesses that make violations impossible.
- Evidence over Plausibility β If it isn't verified by an execution command or physical log, it remains a hypothesis (
[HYPO]). - Stability over Complexity β Fewer moving parts, zero spurious drift, and architectures designed to quietly endure.
- Document along the Way β Systems outlive their builders; code without architectural context is technical debt.
π»π³ Xem bαΊ±ng TiαΊΏng Viα»t (README.vi.md) Β |Β π§ Engineering & Learning Roadmap (Learning.md) Β |Β β Buy Me a Coffee