Applied AI Engineer · Self-hosted LLM and agent systems. I build the memory, context and tuning that make local LLM agents run reliably, and the x402 plumbing that lets them pay for what they need. Most of it on hardware I own, and the rest by choice.
I run a small self-hosted fleet as a lab for two things: making agents work well: the memory, retrieval, context, and model tuning that keep local LLMs reliable over long runs, and letting agents pay for things via x402 (a WordPress plugin, a self-hosted endpoint and a CLI that tallies settlements, so an agent can buy a file or an answer in USDC, self-custodial). Most of it runs on hardware I own, and the parts that do not are a deliberate choice: my seven-role agent system puts four agents on a self-hosted 27B model and three on a hosted frontier model, split by whether the task rewards reasoning or throughput. The thesis behind it all: know when to own, know when to rent.
- 🧠 Context & agent engineering: memory systems, hybrid retrieval, and harness/prompt design that make local LLMs remember more, drift less, and run better. Tuning and hardening setups like Hermes to run reliably over long sessions (hermes-fusion-memory, agent skills).
- 💸 Agent-native commerce (x402): implementations of the x402 protocol, a WordPress plugin, a self-hosted endpoint and a CLI that tallies settlements for accounting. 0% middleman fees, self-custodial, no hosted infra. One real Base mainnet settlement (x402-sandbox, x402-wordpress, x402-books). The protocol is new enough that shipping against it at all is the interesting part; the volume is not the claim.
- 🧩 Agent skills & tooling: reusable systems that give an agent a place to think, remember, plan, and operate: knowledge vaults, resumable task graphs, hybrid memory, and a Kubernetes plugin. See the Skills Built section of my site.
- 🤖 Multi-agent systems: orchestrated CEO / CTO / Auditor-style agent teams (inky-os) that carry a brief through research, production, scheduling and invoicing. Agents propose, humans approve: every generated asset passes a gate before release. Open-source reference implementation with the autonomy boundary documented, AGPL.
- 🎬 "The Press" content engine: programmatic video reels (Remotion, 15+ styles), carousels, branded PDFs, and TTS voiceovers, generated end-to-end from a campaign brief.
- 🌍 Production web estate: a cluster of WordPress + WooCommerce storefronts and a self-hosted Moodle course platform with an onboard tutor LLM, containerized behind Caddy + Cloudflare Tunnels across multiple hosts.
realandworks.com: case-study builds, the skills I've shipped, open-source tools, and my full stack. Built and self-hosted end to end.
| Project | What it is |
|---|---|
| inky-os | Multi-agent "company OS", a CEO + specialist agents that take a brief through research, production, scheduling & invoicing, with a human approval gate before anything ships. File-first memory, immutable logging, hybrid local + hosted inference (AGPL) |
| x402-sandbox | Self-hosted x402 endpoint, make any file, folder or endpoint payable by an AI agent in USDC. Demand pricing, sales funnel, admin dashboard |
| x402-wordpress | Sell files, endpoints, and answers to AI agents for USDC over x402, self-custodial WordPress plugin, 0% fees, no middleman |
| zhc-fabric | Multi-view consensus fabric for agent teams, an Erlang/OTP sidecar in Docker (no host Erlang) |
| hermes-fusion-memory | Hybrid agent memory, Fusion Brain + MemPalace: per-turn recall, living playbooks, a memory_search tool |
| ATG-looping | Atomic Task Graph agent skills for Claude Code, graph-driven planning + resumable looping (built from arXiv:2607.01942) |
| moodle-tool_h5pmedic | Upgrade H5P Medic, make Moodle upgrades safe: fingerprint, compare, and auto-repair H5P (GPL · PHPUnit + CI) |
- Languages: Python, PHP, TypeScript / JavaScript, Bash, SQL (a little Erlang/OTP)
- Infra: Docker · Docker Compose · Caddy · Cloudflare Tunnels · systemd · Linux (Ubuntu & Arch)
- Currently learning: Kubernetes (six-month hands-on course from Jul 2026, no cert yet; the deliverable is a running homelab cluster, which I will publish rather than claim)
- AI/ML: llama.cpp · local LLMs (Qwen) · Claude Code & Agent Skills · MCP · RAG · memory systems · n8n · Remotion
- Agent payments: x402 · USDC on Base · self-custodial CDP
- Data / CMS: MariaDB · PostgreSQL · WordPress / WooCommerce · Moodle
- CRM / Automation: GoHighLevel
2026, Self-hosted AI & agent engineering. Context & memory systems that keep local LLM agents reliable, x402 payment plumbing, human-gated agent teams, a content engine, and a live WordPress/WooCommerce + Moodle estate, across a multi-machine fleet I own, with hosted models where they earn their place.
2025, Standing start. Came into AI engineering from zero in early 2025 and started shipping production systems the same year. The pace is the point: I learn a stack by putting it into live use.
Before that. Banking, then three years building interactive e-learning software (shipped on a PC Magazine cover DVD), then two decades away from code: property, photography, digital marketing, and regulated fintech client onboarding. One client has stayed with me since 2015.
- Portfolio: realandworks.com
- X: @InkyPyrus_PubCo
- LinkedIn: Regard Vermeulen
Always learning. Thanks for visiting! 🌟