AI Infrastructure Architect building PAI — a Personal AI Infrastructure system that turns Claude Code into a persistent, self-improving development environment.
PAI is an open architecture that wraps Claude Code with skills, hooks, memory, and a continuously upgrading algorithm. It's not a chatbot — it's scaffolding that makes AI reliable, repeatable, and personal.
Built on Daniel Miessler's Personal AI Infrastructure framework and Fabric pattern system.
Current state (v4.0.3):
| Component | Count | What it does |
|---|---|---|
| Skills | 86 | Self-activating domain expertise — security, research, creative writing, OSINT, video production, and more |
| Hooks | 37 | Event lifecycle handlers — session start/stop, tool validation, memory capture, security scanning |
| Algorithm | v3.7.0 | 7-phase execution engine (Observe → Think → Plan → Build → Execute → Verify → Learn) |
| Memory | 16,700+ sessions | SQLite + FTS5 + embeddings — persistent context across every conversation |
| Agents | 14 types | Specialized sub-agents for engineering, architecture, research, security, design |
| Fabric Patterns | 240+ | Content analysis, extraction, and transformation templates |
- Scaffolding > Model — Architecture matters more than which LLM you use
- Code Before Prompts — If code can solve it, don't prompt for it
- As Deterministic As Possible — Same input, same output, always
- The Algorithm Is The Centerpiece — Everything else exists to feed Current State → Ideal State
- Memory Makes Intelligence Compound — Without persistence, every session starts from zero
Every non-trivial task runs through a 7-phase loop that transitions from Current State to Ideal State via verifiable criteria:
OBSERVE → THINK → PLAN → BUILD → EXECUTE → VERIFY → LEARN
↑ │
└────────────────────────────────────────────────────┘
Each phase has explicit gates. The algorithm self-improves from accumulated evidence across sessions.
| Component | Details |
|---|---|
| Hosts | Multi-node homelab — Docker, K3s, Incus, Tailscale mesh |
| GPUs | NVIDIA GPU compute (CUDA, vLLM inference) |
| Networking | Tailscale overlay, Traefik reverse proxy, auto-TLS |
| Observability | Prometheus, Grafana, structured logging |
| Backups | Restic to B2, systemd timers, K3s + Incus state included |
- AI/ML: Claude Code, vLLM, RAG pipelines, embeddings, NVIDIA CUDA/DCGM
- Runtime: TypeScript/Bun, Python, Bash
- Infra: Docker, Kubernetes (K3s), Incus containers, Cloudflare Workers
- Ops: Prometheus/Grafana, systemd, Traefik, Tailscale, restic
Complexity is borrowed — every layer added is future time invested.
Record your dead ends — failed approaches prevent wasted future effort.
Silent failures are worst — if it can fail, make it fail loud.
Spec/Test/Evals first — if you can't specify it, you can't trust it.
- GitHub: @nixfred
- LinkedIn: nixfred
- Site: nixfred.com
Building compounding AI infrastructure, one session at a time.