I build production systems where the orchestration is the product — multi-agent LLM pipelines, domain-expert Claude Code plugins, and autonomous trading infrastructure. Most of my recent work has been turning tribal knowledge into inspectable, eval-scored, agent-addressable assets.
Current focus — autonomous futures trading system (paper-validated on MCL, MNQ, MES) with a 7-layer regime classifier, Kelly-sized position management, and a full Sonnet-executor / Opus-grader evaluation harness.
Currently exploring Solutions Architect roles where the architecture decisions are driven by domain reality, not vendor catalogs.
| Repo | What it is | Live |
|---|---|---|
| ralph-orchestrator | Reusable primitives for multi-agent LLM systems — circuit breakers, token tracking, session memory, health metrics, and a drop-in FastAPI router. Extracted from a working production deployment. 50/50 tests, CI green on Python 3.10/3.11/3.12. | pages |
| dp600-stark-v2 | DP-600 Microsoft Fabric Analytics Engineer exam prep — offline-first PWA with 1,287 questions, 302 SM-2-lite flashcards, 60 scenario chains. 258 tests pass. | live |
- Eval-driven development — every skill, every model call, every strategy has a failing test first.
- Progressive disclosure — skills load only when relevant; references load only when deep-linked. Context is the scarcest resource in an agent system.
- Safety rails inside the artifact, not around it — if a skill can move money or data, its first section says so. Preview, diff, confirm.
- Extraction over invention — the best libraries are the ones carved out of working code, not designed in the abstract.
GitHub issues on any of my repos; LinkedIn; mattamundson@greenfieldmetalsales.com.