Adrian Wedd — systems architect (Tasmania)

I build evaluation infrastructure for agentic AI and ship the surrounding tooling — MCP servers, voice and embodied interfaces, ops plumbing. Seven years at Tasmania's Department of Communities; wrote Homes Tasmania's first generative-AI policy. No formal degree.

Now

• failure-first-embodied-ai — adversarial evaluation framework for agentic AI. 13,988 scenarios across 414 attack classes; multi-model vulnerability runs.
• spark — Claude-powered companion robot on PiCar-X + Raspberry Pi. Built for a neurodivergent kid; declarative-language interaction, deterministic crisis fallbacks.
• afterwords — zero-shot voice cloning on Apple Silicon. Gives Claude Code a voice. ~15 voices included.
• rlm-mcp — MCP server implementing the Recursive Language Model pattern. Lets Claude Code work over 1M+ character documents via session-based chunking and BM25.

Shipped

• thiswasntinthebrochure.wtf — co-parenting ADHD, autism, PDA & ODD. Four localised editions (AU/UK/US/NZ), ≥70% peer-reviewed citations.
• Agentic-Index — scored, curated catalogue of autonomous AI tooling.
• adrianwedd.com — living workshop. AI safety writing, NotebookLM-generated audio + cinematic video on every post, Astro 6 + Tailwind 4.
• failure-first — site for the evaluation work.
• Client work: Evolve Chiropractic, Wolf Clan Zen Do Kai, Tassie Mini-Z, Clutterbusters, Readings with Rowena.

Stack

Python, TypeScript, Bash. Anthropic SDK, MCP, LangGraph, OpenRouter, vLLM, Ollama. FastAPI, Astro, Cloudflare Workers + R2, PostgreSQL, Redis. Pytest, Playwright, custom adversarial harnesses.

Links

adrianwedd.com · failurefirst.org · thiswasntinthebrochure.wtf · status