Portfolio · 2026

Anthony Windey

Agentic AI & IBM i developer

This year I designed a harness that lets Claude Code develop safely, an AI-assisted development setup for IBM i, and a self-hosted multi-agent assistant that I run. Every number below is measured.

43 · 23skills · hooks I designed for Claude Code
7,836commits in 2026 over 5 repositories, all written by Claude Code
1,880notes in the project vault (its knowledge base), 482 of them decisions
2 × RTX 5090self-hosted inference, image and video generation, local workers for Claude Code

Claude Code harness

Plugin of skills and hooks · session CLI · MCP servers · local workers

/analyze/dev/critic

dc landdc deploy/vault-sync

Claude Codepi × NQwen

a hook blocks the next step worker citations checked in the code

Method as skills. /analyze plans, /dev implements, /critic checks the diff, dc (an in-house tool) merges and deploys, /vault-sync records the lessons.
MRS Memory. A wiki-style memory engine with hybrid search: embeddings + BM25, RRF fusion, cross-encoder rerank; recall tested on golden sets.
Enforced by hooks. No edit before the plan, no merge before the review; one git worktree per session.
Local workers. Big reading jobs are split into one-question tasks for local agents (pi) on Qwen, at no API cost.

AI-assisted development on IBM i

Agent in VS Code on the workstation, reaching the IBM i over SSH · MCP

AGENT ENTRY · IFS SCRIPTNO SHELL

1 Browse4 Check out2 Create version5 Update source3 Create component6 Compile -> feedback

Selection ===>

Illustration The whole change cycle, behind one SSH key

Map first. An in-house skill documents each business concept a task touches and guides navigation to the code.
Context. IBM i sources exported to the IFS; project documentation synced to the workstation.
One script, full loop. It runs every step through compilation, ARCAD versions and check-outs included, and returns the errors; the agent fixes until it compiles. 5250 green screens driven through MCP, mirrored live in a browser.

MRS Studio

Self-hosted multi-agent assistant · designed and directed alone · since March

Web researchMemoryEmailCalendarMain
assistantDocumentsMediaWorkflowsSettings

Qwen on 2 local GPUs · Claude if chosen

One specialist per domain. Each request goes to one of eight specialists, each with its own tools and context.
A brain that consolidates. The assistant extracts facts from conversations and mail; a consolidation step keeps the lasting ones and flags duplicates.
On my own hardware. Quantized Qwen models (vLLM, NInfer, Strata) behind an OpenAI-compatible gateway.

Lab

Work on local inference